Improvement study of an aluminum sulfate production plant
Authorship
N.A.L.
Master's Degree in Chemical Engineering and Bioprocess
N.A.L.
Master's Degree in Chemical Engineering and Bioprocess
Defense date
02.19.2026 12:45
02.19.2026 12:45
Summary
The main objective of this study is the techno-economic evaluation of a modification to the manufacturing process of aluminum sulfate solution. The analysis focuses on replacing the main reactive raw material as a strategy to improve operational performance, optimize resource use, and ensure the sustainability of the process, while maintaining the quality of the final product and compliance with all industrial safety requirements. Based on a review of the conventional bauxite-based system, opportunities for improvement are identified related to the complexity of treating impure raw materials and managing solid waste. In this context, the use of high-purity aluminum hydroxide is proposed and analyzed in detail as an alternative. To support the proposal, this report develops a detailed description of the modified process, including mass and energy balances for different production scenarios, operational planning using Gantt charts, and a preliminary economic evaluation that compares the costs of raw materials, waste treatment, and equipment amortization for both routes.
The main objective of this study is the techno-economic evaluation of a modification to the manufacturing process of aluminum sulfate solution. The analysis focuses on replacing the main reactive raw material as a strategy to improve operational performance, optimize resource use, and ensure the sustainability of the process, while maintaining the quality of the final product and compliance with all industrial safety requirements. Based on a review of the conventional bauxite-based system, opportunities for improvement are identified related to the complexity of treating impure raw materials and managing solid waste. In this context, the use of high-purity aluminum hydroxide is proposed and analyzed in detail as an alternative. To support the proposal, this report develops a detailed description of the modified process, including mass and energy balances for different production scenarios, operational planning using Gantt charts, and a preliminary economic evaluation that compares the costs of raw materials, waste treatment, and equipment amortization for both routes.
Direction
GARRIDO FERNANDEZ, JUAN MANUEL (Tutorships)
GARRIDO FERNANDEZ, JUAN MANUEL (Tutorships)
Court
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
Valorization of macroalgae for the recovery of bioactive compounds through the cultivation of purple phototrophic bacteria (PPB)
Authorship
M.A.G.
Master's Degree in Chemical Engineering and Bioprocess
M.A.G.
Master's Degree in Chemical Engineering and Bioprocess
Defense date
02.19.2026 11:15
02.19.2026 11:15
Summary
Macroalgal accumulations on the Galician coast caused by strandings have raised concerns about their economic and environmental impacts. This Master’s Thesis evaluates two parallel processes aimed at valorizing this natural residue and recovering value-added products such as carotenoids. After milling, the algae at 5 g COD per liter were first subjected to a prefermentation using sludge from an anaerobic sludge digester in one reactor and another with purple phototrophic bacteria (PPB) under dark conditions. The sludge reactor consumed 31 percent of the COD and accumulated 1,38 grams of COD per liter of volatile fatty acids VFA, mainly acetic and butyric acid. The PPB reactor reached 1,68 grams of COD per liter of VFA, with propionic acid prevailing. In parallel, a PPB reactor operated under infrared light with 0,5 grams of COD per liter of algae to assess its potential for direct transformation. Its performance was limited by algal hydrolysis, and biomass interfered with carotenoid quantification due to a large percentage of macroalgal solids remaining throughout the assay. The supernatants from both prefermentation reactors were fed into a second phototrophic PPB stage, yielding 0,664 and 1,164 milligrams of betacarotene per gram of dry algae consumed. These results demonstrate the metabolic versatility of PPB and provide a circular option for managing algal strandings. However, more research is needed to optimize the obtained yields.
Macroalgal accumulations on the Galician coast caused by strandings have raised concerns about their economic and environmental impacts. This Master’s Thesis evaluates two parallel processes aimed at valorizing this natural residue and recovering value-added products such as carotenoids. After milling, the algae at 5 g COD per liter were first subjected to a prefermentation using sludge from an anaerobic sludge digester in one reactor and another with purple phototrophic bacteria (PPB) under dark conditions. The sludge reactor consumed 31 percent of the COD and accumulated 1,38 grams of COD per liter of volatile fatty acids VFA, mainly acetic and butyric acid. The PPB reactor reached 1,68 grams of COD per liter of VFA, with propionic acid prevailing. In parallel, a PPB reactor operated under infrared light with 0,5 grams of COD per liter of algae to assess its potential for direct transformation. Its performance was limited by algal hydrolysis, and biomass interfered with carotenoid quantification due to a large percentage of macroalgal solids remaining throughout the assay. The supernatants from both prefermentation reactors were fed into a second phototrophic PPB stage, yielding 0,664 and 1,164 milligrams of betacarotene per gram of dry algae consumed. These results demonstrate the metabolic versatility of PPB and provide a circular option for managing algal strandings. However, more research is needed to optimize the obtained yields.
Direction
Pedrouso Fuentes, Alba (Tutorships)
MOSQUERA CORRAL, ANUSKA (Co-tutorships)
Pedrouso Fuentes, Alba (Tutorships)
MOSQUERA CORRAL, ANUSKA (Co-tutorships)
Court
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
Explicit Thing/Stuff Conditioning for Robust Open-Vocabulary Panoptic Segmentation
Authorship
A.B.R.
Master's Degree in Artificial Intelligence
A.B.R.
Master's Degree in Artificial Intelligence
Defense date
02.03.2026 13:30
02.03.2026 13:30
Summary
Open-vocabulary panoptic segmentation aims to segment and classify arbitrary categories at test time using language priors, but current methods still struggle under vocabulary shift, especially with ambiguous thing/stuff distinction: unseen categories may be incorrectly treated as stuff instead of thing, causing instance masks to be erroneously merged, hurting panoptic quality. We propose TS-CLIP, a lightweight model that makes grouping explicit by conditioning CLIP text prototypes on the desired panoptic type (thing/stuff) and global image context, and by adding query-to-text attention so category semantics can influence region formation. A training-time fusion/split augmentation further prevents the model from memorizing dataset-specific grouping conventions and allows faster training convergence. With additional decoder tweaks, TS-CLIP improves over FC-CLIP on the ADE20K dataset (PQ from 26.8 to 27.6), particularly on thing (PQ from 25.1 to 26.5) and unseen categories (PQ from 18.0 to 19.1), while taking fewer iterations to train.
Open-vocabulary panoptic segmentation aims to segment and classify arbitrary categories at test time using language priors, but current methods still struggle under vocabulary shift, especially with ambiguous thing/stuff distinction: unseen categories may be incorrectly treated as stuff instead of thing, causing instance masks to be erroneously merged, hurting panoptic quality. We propose TS-CLIP, a lightweight model that makes grouping explicit by conditioning CLIP text prototypes on the desired panoptic type (thing/stuff) and global image context, and by adding query-to-text attention so category semantics can influence region formation. A training-time fusion/split augmentation further prevents the model from memorizing dataset-specific grouping conventions and allows faster training convergence. With additional decoder tweaks, TS-CLIP improves over FC-CLIP on the ADE20K dataset (PQ from 26.8 to 27.6), particularly on thing (PQ from 25.1 to 26.5) and unseen categories (PQ from 18.0 to 19.1), while taking fewer iterations to train.
Direction
MUCIENTES MOLINA, MANUEL FELIPE (Tutorships)
CORES COSTA, DANIEL (Co-tutorships)
MUCIENTES MOLINA, MANUEL FELIPE (Tutorships)
CORES COSTA, DANIEL (Co-tutorships)
Court
BUGARIN DIZ, ALBERTO JOSE (Chairman)
VALLADARES RODRIGUEZ, SONIA MARIA (Secretary)
PICHEL CAMPOS, JOSE RAMON (Member)
BUGARIN DIZ, ALBERTO JOSE (Chairman)
VALLADARES RODRIGUEZ, SONIA MARIA (Secretary)
PICHEL CAMPOS, JOSE RAMON (Member)
Fatty acid-based eutectic solvents for the improvement of gelatine extraction process from fish skin.
Authorship
M.D.A.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
M.D.A.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
Defense date
02.19.2026 11:45
02.19.2026 11:45
Summary
Playing a central role in the Galician economy, the fish processing industry generates significant volumes of waste that, despite their current low exploitation, exhibit a high valorization potential. Among them, fish skin stands out for its high collagen content, which can be used as a raw material for the production of gelatin, with multiple applications across diverse sectors. Conventional gelatin extraction processes employ pretreatments with strong acids and bases under aggressive conditions (high concentrations or prolonged times), which, in addition to compromising the process sustainability credentials, induce uncontrolled hydrolysis that substantially impairs the quality of the final biopolymer. Furthermore, these methodologies require intensive neutralization stages and involve environmental risks associated with the handling and discharge of the generated corrosive effluents. In this line, it is of interest to investigate more sustainable alternative methods. Recently, the production of gelatin from fish skin using warm-water extraction following a maceration pretreatment with eutectic solvents has been proposed. In this context, this work investigates the application of eutectic solvents based on quaternary ammonium salts and fatty acids. First, the solid-liquid equilibrium of eight binary systems constituted by ammonium salt (tetraethylammonium chloride or tetrapropylammonium chloride) and saturated fatty acid (octanoic, decanoic, dodecanoic, or octadecanoic) was analyzed using differential scanning calorimetry. Based on this screening and considering the measured eutectic behavior of the aforementioned systems, the two eutectic solvents containing octanoic acid were selected, as they possessed the lowest eutectic temperatures (below 0 Celsius). Subsequently, the extraction process was evaluated, considering the presence or absence of a cold-water washing step between the eutectic solvent pretreatment and the final water extraction. The obtained gelatins were characterized by thermal stability analysis, molecular weight distribution, and rheological behavior. The results indicated extraction yields of 6-10%, as well as properties suggesting the production of a high-quality product: gelling and melting temperatures of 24-27 Celsius and 27-29 Celsius, respectively; decomposition temperatures above 225 Celsius; and a clear predominance of high-molecular-weight fractions. Among the evaluated options, the most promising results were obtained using the tetrapropylammonium chloride and octanoic acid eutectic solvent during the maceration step with the intermediate washing step. In conclusion, this work demonstrates the technical feasibility of using eutectic solvents composed of quaternary ammonium salts and fatty acids to produce high-quality fish skin gelatin with promising application prospects.
Playing a central role in the Galician economy, the fish processing industry generates significant volumes of waste that, despite their current low exploitation, exhibit a high valorization potential. Among them, fish skin stands out for its high collagen content, which can be used as a raw material for the production of gelatin, with multiple applications across diverse sectors. Conventional gelatin extraction processes employ pretreatments with strong acids and bases under aggressive conditions (high concentrations or prolonged times), which, in addition to compromising the process sustainability credentials, induce uncontrolled hydrolysis that substantially impairs the quality of the final biopolymer. Furthermore, these methodologies require intensive neutralization stages and involve environmental risks associated with the handling and discharge of the generated corrosive effluents. In this line, it is of interest to investigate more sustainable alternative methods. Recently, the production of gelatin from fish skin using warm-water extraction following a maceration pretreatment with eutectic solvents has been proposed. In this context, this work investigates the application of eutectic solvents based on quaternary ammonium salts and fatty acids. First, the solid-liquid equilibrium of eight binary systems constituted by ammonium salt (tetraethylammonium chloride or tetrapropylammonium chloride) and saturated fatty acid (octanoic, decanoic, dodecanoic, or octadecanoic) was analyzed using differential scanning calorimetry. Based on this screening and considering the measured eutectic behavior of the aforementioned systems, the two eutectic solvents containing octanoic acid were selected, as they possessed the lowest eutectic temperatures (below 0 Celsius). Subsequently, the extraction process was evaluated, considering the presence or absence of a cold-water washing step between the eutectic solvent pretreatment and the final water extraction. The obtained gelatins were characterized by thermal stability analysis, molecular weight distribution, and rheological behavior. The results indicated extraction yields of 6-10%, as well as properties suggesting the production of a high-quality product: gelling and melting temperatures of 24-27 Celsius and 27-29 Celsius, respectively; decomposition temperatures above 225 Celsius; and a clear predominance of high-molecular-weight fractions. Among the evaluated options, the most promising results were obtained using the tetrapropylammonium chloride and octanoic acid eutectic solvent during the maceration step with the intermediate washing step. In conclusion, this work demonstrates the technical feasibility of using eutectic solvents composed of quaternary ammonium salts and fatty acids to produce high-quality fish skin gelatin with promising application prospects.
Direction
RODRIGUEZ MARTINEZ, HECTOR (Tutorships)
MENDES VILAS BOAS, SERGIO ANTONIO (Co-tutorships)
RODRIGUEZ MARTINEZ, HECTOR (Tutorships)
MENDES VILAS BOAS, SERGIO ANTONIO (Co-tutorships)
Court
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
Design and Comparative Evaluation of Advanced Safeguard Nodes for Conversational AI
Authorship
A.E.C.
Master's Degree in Artificial Intelligence
A.E.C.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 16:30
02.18.2026 16:30
Summary
This paper presents the design and evaluation of an efficient 'Safeguard Node' to protect conversational Artificial Intelligence (AI) systems in production environments. The author proposes a Unified Multi-Head architecture based on Bidirectional Encoder Representations from Transformers (BERT) encoders, ranging from approximately 4M to 110M parameters, that simultaneously predicts semantic category, user intent, and risk level. The study addresses the high computational costs and latency associated with current safety models, such as Llama Guard. To this end, a trilingual dataset (English, Spanish, and Galician) was developed, covering nine semantic categories, two intent classes, and two risk levels. The results demonstrate that these smaller, specialized encoders can match or outperform much larger Large Language Models (LLMs) with 4B or 7B parameters in specific moderation tasks, drastically reducing latency, hardware requirements, and costs associated with external Application Programming Interfaces (APIs).
This paper presents the design and evaluation of an efficient 'Safeguard Node' to protect conversational Artificial Intelligence (AI) systems in production environments. The author proposes a Unified Multi-Head architecture based on Bidirectional Encoder Representations from Transformers (BERT) encoders, ranging from approximately 4M to 110M parameters, that simultaneously predicts semantic category, user intent, and risk level. The study addresses the high computational costs and latency associated with current safety models, such as Llama Guard. To this end, a trilingual dataset (English, Spanish, and Galician) was developed, covering nine semantic categories, two intent classes, and two risk levels. The results demonstrate that these smaller, specialized encoders can match or outperform much larger Large Language Models (LLMs) with 4B or 7B parameters in specific moderation tasks, drastically reducing latency, hardware requirements, and costs associated with external Application Programming Interfaces (APIs).
Direction
CATALA BOLOS, ALEJANDRO (Tutorships)
Piñeiro Martín, Andrés (Co-tutorships)
CATALA BOLOS, ALEJANDRO (Tutorships)
Piñeiro Martín, Andrés (Co-tutorships)
Court
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
A Human-in-the-Loop Study on Prompt Optimization and Model Comparison for Non-Functional Requirements Classification
Authorship
R.U.F.
Master's Degree in Artificial Intelligence
R.U.F.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 18:00
02.18.2026 18:00
Summary
Non-functional requirements (NFRs) are critical factors in Requirements Engineering, as neglecting them results in significant problems. To adequately address them, automated classification of NFRs into subclasses is a current research area. Traditionally, supervised machine learning models were used for this task. However, there is a severe scarcity of labeled data for languages other than English, which is required for effective training. This work compares encoder-only transformer models to generative Large Language Models (LLMs) in a multi-class NFR classification task. Specifically, it is investigated whether humanin-the-loop prompt optimization can achieve results competitive to the fine-tuned transformer models. To achieve these insights, the PROMISE translated dataset is used to train the model and optimize the prompt in a 3-fold-stratified cross-validation setup, and the ReSpaN dataset to evaluate the cross-dataset generalization of the models. Performance is measured via macro F1-score, to carefully handle class imbalances. The findings of this study demonstrate that generative LLMs of reasonable size (7B+ parameters) outperformed fine-tuned transformer models in this low-data scenario. The optimization of the prompt seems to overfit on the training data initially, but demonstrates generalization when tested on the ReSpaN dataset. Furthermore, the performance of larger generative LLMs remains stable across English and Spanish data. In conclusion, generative LLMs provide a robust solution for NFR classification because they are not dependent on large labeled datasets and can reason across language barriers. By reducing the data dependency, this approach simplifies the automation of Requirements Engineering in languages other than English.
Non-functional requirements (NFRs) are critical factors in Requirements Engineering, as neglecting them results in significant problems. To adequately address them, automated classification of NFRs into subclasses is a current research area. Traditionally, supervised machine learning models were used for this task. However, there is a severe scarcity of labeled data for languages other than English, which is required for effective training. This work compares encoder-only transformer models to generative Large Language Models (LLMs) in a multi-class NFR classification task. Specifically, it is investigated whether humanin-the-loop prompt optimization can achieve results competitive to the fine-tuned transformer models. To achieve these insights, the PROMISE translated dataset is used to train the model and optimize the prompt in a 3-fold-stratified cross-validation setup, and the ReSpaN dataset to evaluate the cross-dataset generalization of the models. Performance is measured via macro F1-score, to carefully handle class imbalances. The findings of this study demonstrate that generative LLMs of reasonable size (7B+ parameters) outperformed fine-tuned transformer models in this low-data scenario. The optimization of the prompt seems to overfit on the training data initially, but demonstrates generalization when tested on the ReSpaN dataset. Furthermore, the performance of larger generative LLMs remains stable across English and Spanish data. In conclusion, generative LLMs provide a robust solution for NFR classification because they are not dependent on large labeled datasets and can reason across language barriers. By reducing the data dependency, this approach simplifies the automation of Requirements Engineering in languages other than English.
Direction
CONDORI FERNANDEZ, OLINDA NELLY (Tutorships)
CATALA BOLOS, ALEJANDRO (Co-tutorships)
CONDORI FERNANDEZ, OLINDA NELLY (Tutorships)
CATALA BOLOS, ALEJANDRO (Co-tutorships)
Court
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
Anchoring catalytic nanoparticles to polyvinylidene fluoride fibers with an encapsulated phase change material for thermo-self-regulation
Authorship
A.F.M.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
A.F.M.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
Defense date
02.19.2026 12:15
02.19.2026 12:15
Summary
This Master’s Thesis presents a proof of concept for the integration of PVDF-based polymer fibers with encapsulated phase change materials (PCM) and catalytic nanoparticles, aiming to develop a thermally self-regulating catalytic system. The fibers were fabricated using a coaxial microfluidic non-solvent induced phase separation (NIPS) process, enabling controlled encapsulation of RT25HC paraffin with different PCM loadings. Surface functionalization was achieved by spray coating iron oxide (Fe3O4) nanoparticles, selected after a preliminary colloidal stability study, and their adhesion, stability, and compatibility with the thermal behavior of the fibers were systematically evaluated.
This Master’s Thesis presents a proof of concept for the integration of PVDF-based polymer fibers with encapsulated phase change materials (PCM) and catalytic nanoparticles, aiming to develop a thermally self-regulating catalytic system. The fibers were fabricated using a coaxial microfluidic non-solvent induced phase separation (NIPS) process, enabling controlled encapsulation of RT25HC paraffin with different PCM loadings. Surface functionalization was achieved by spray coating iron oxide (Fe3O4) nanoparticles, selected after a preliminary colloidal stability study, and their adhesion, stability, and compatibility with the thermal behavior of the fibers were systematically evaluated.
Direction
RODRIGUEZ MARTINEZ, HECTOR (Tutorships)
Durán López, Mikel (Co-tutorships)
RODRIGUEZ MARTINEZ, HECTOR (Tutorships)
Durán López, Mikel (Co-tutorships)
Court
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
MOREIRA VILAR, MARIA TERESA (Chairman)
RODIL RODRIGUEZ, EVA (Secretary)
PIÑEIRO CHOUSA, JUAN RAMON (Member)
A Multimodal Real-Time Cross-Language Speech Emotion Recognition System
Authorship
S.A.F.V.
Master's Degree in Artificial Intelligence
S.A.F.V.
Master's Degree in Artificial Intelligence
Defense date
02.03.2026 14:00
02.03.2026 14:00
Summary
Real-Time Speech Emotion Recognition (SER) in multilingual environments remains a significant challenge due to the semantic gap between acoustic and linguistic cues and the computational cost of modern deep learning models. This paper presents a multimodal, cross-language SER system designed to balance predictive performance with environmental efficiency on Real-Time scenarios. We proposed a hybrid architecture that integrates acoustic features (MFCCs) and semantic embeddings (DistilUse) via diverse fusion strategies, ranging from concatenation to cross-attention mechanisms. To address the environmental dimension of ethical AI, we integrate a Green AI monitoring layer using CodeCarbon to quantify energy consumption during training and inference. Experimental validation was conducted on the multilingual EmoFilm dataset (English, Spanish, Italian). Results indicate that a Late Fusion approach (alpha equals to 0.5) outperforms complex Mid- Fusion architectures, achieving an accuracy of 70.19 percent, likely due to the limited dataset size. Furthermore, a cross-lingual analysis reveals a possible dubbing effect (i.e., improved recognition performance on professionally dubbed audio, where exaggerated prosody may amplify emotional cues), where the model performs significantly better on dubbed audio (Spanish, 77.3 percent) compared to original tracks (English, 65.0 percent), due to accentuated prosody in voice acting. Real-time experiments demonstrate that the proposed dual-rate asynchronous architecture operates well below real-time constraints (RTF less than 1) maintaining meaningful recognition performance under streaming conditions. Carbon footprint analysis highlights that lightweight acoustic representations combined with semantic processing provide a favourable accuracy-efficiency trade-off. Overall, the results indicate that effective multimodal SER in real-time multilingual settings does not require complex models, but rather carefully designed, efficiency-aware system architectures.
Real-Time Speech Emotion Recognition (SER) in multilingual environments remains a significant challenge due to the semantic gap between acoustic and linguistic cues and the computational cost of modern deep learning models. This paper presents a multimodal, cross-language SER system designed to balance predictive performance with environmental efficiency on Real-Time scenarios. We proposed a hybrid architecture that integrates acoustic features (MFCCs) and semantic embeddings (DistilUse) via diverse fusion strategies, ranging from concatenation to cross-attention mechanisms. To address the environmental dimension of ethical AI, we integrate a Green AI monitoring layer using CodeCarbon to quantify energy consumption during training and inference. Experimental validation was conducted on the multilingual EmoFilm dataset (English, Spanish, Italian). Results indicate that a Late Fusion approach (alpha equals to 0.5) outperforms complex Mid- Fusion architectures, achieving an accuracy of 70.19 percent, likely due to the limited dataset size. Furthermore, a cross-lingual analysis reveals a possible dubbing effect (i.e., improved recognition performance on professionally dubbed audio, where exaggerated prosody may amplify emotional cues), where the model performs significantly better on dubbed audio (Spanish, 77.3 percent) compared to original tracks (English, 65.0 percent), due to accentuated prosody in voice acting. Real-time experiments demonstrate that the proposed dual-rate asynchronous architecture operates well below real-time constraints (RTF less than 1) maintaining meaningful recognition performance under streaming conditions. Carbon footprint analysis highlights that lightweight acoustic representations combined with semantic processing provide a favourable accuracy-efficiency trade-off. Overall, the results indicate that effective multimodal SER in real-time multilingual settings does not require complex models, but rather carefully designed, efficiency-aware system architectures.
Direction
CONDORI FERNANDEZ, OLINDA NELLY (Tutorships)
CATALA BOLOS, ALEJANDRO (Co-tutorships)
CONDORI FERNANDEZ, OLINDA NELLY (Tutorships)
CATALA BOLOS, ALEJANDRO (Co-tutorships)
Court
BUGARIN DIZ, ALBERTO JOSE (Chairman)
VALLADARES RODRIGUEZ, SONIA MARIA (Secretary)
PICHEL CAMPOS, JOSE RAMON (Member)
BUGARIN DIZ, ALBERTO JOSE (Chairman)
VALLADARES RODRIGUEZ, SONIA MARIA (Secretary)
PICHEL CAMPOS, JOSE RAMON (Member)
Interactive Web System for Population Medical Data Analysis
Authorship
S.G.
Master's Degree in Computer Vision
S.G.
Master's Degree in Computer Vision
Defense date
02.04.2026 10:30
02.04.2026 10:30
Summary
This work presents a web-based platform for the visualization, analysis, and segmentation of medical imaging and structured clinical data. The system integrates multiple modules, including statistical analysis, interactive image viewing, image management, and automated segmentation, to support exploratory research workflows. The experimental results confirm the correct operation of all modules and demonstrate the platform’s ability to handle heterogeneous medical data in an interactive and user-friendly manner. The overall design is modular and flexible, allowing the system to be extended and adapted for future research and potential clinical use.
This work presents a web-based platform for the visualization, analysis, and segmentation of medical imaging and structured clinical data. The system integrates multiple modules, including statistical analysis, interactive image viewing, image management, and automated segmentation, to support exploratory research workflows. The experimental results confirm the correct operation of all modules and demonstrate the platform’s ability to handle heterogeneous medical data in an interactive and user-friendly manner. The overall design is modular and flexible, allowing the system to be extended and adapted for future research and potential clinical use.
Direction
NUÑEZ GARCIA, MARTA (Tutorships)
NUÑEZ GARCIA, MARTA (Tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
Few-Shot Segmentation for Medical Imaging Using Foundation Models
Authorship
J.M.G.D.
Master's Degree in Computer Vision
J.M.G.D.
Master's Degree in Computer Vision
Defense date
02.04.2026 10:10
02.04.2026 10:10
Summary
Medical image segmentation is a critical prerequisite for diagnosis and treatment planning. While supervised deep learning models have established state-of-the-art performance, they suffer from a heavy reliance on large-scale, pixel-level annotated datasets. This dependency is a significant bottleneck in medical imaging due to the scarcity of expert annotations and the heterogeneity of image modalities. This thesis proposes a novel Few-Shot Segmentation (FSS) framework designed to address these challenges by leveraging Foundation Models (FMs). The proposed method combines the robust feature extraction of the self-supervised DINOv3 with the Segment Anything Model 3 (SAM 3) final boundary refinement. We evaluate this framework across five distinct medical imaging datasets. The experimental results demonstrate that our approach not only generalizes better to unseen classes in low-data scenarios but also surpasses the Dice similarity coefficient of the standard supervised U-Net, marking a significant step forward in label-efficient medical image analysis.
Medical image segmentation is a critical prerequisite for diagnosis and treatment planning. While supervised deep learning models have established state-of-the-art performance, they suffer from a heavy reliance on large-scale, pixel-level annotated datasets. This dependency is a significant bottleneck in medical imaging due to the scarcity of expert annotations and the heterogeneity of image modalities. This thesis proposes a novel Few-Shot Segmentation (FSS) framework designed to address these challenges by leveraging Foundation Models (FMs). The proposed method combines the robust feature extraction of the self-supervised DINOv3 with the Segment Anything Model 3 (SAM 3) final boundary refinement. We evaluate this framework across five distinct medical imaging datasets. The experimental results demonstrate that our approach not only generalizes better to unseen classes in low-data scenarios but also surpasses the Dice similarity coefficient of the standard supervised U-Net, marking a significant step forward in label-efficient medical image analysis.
Direction
VILA BLANCO, NICOLAS (Tutorships)
CORES COSTA, DANIEL (Co-tutorships)
VILA BLANCO, NICOLAS (Tutorships)
CORES COSTA, DANIEL (Co-tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
Adaptation of StyleGAN3 for Processing Multidimensional Remote Sensing Images
Authorship
A.G.L.
Master's Degree in Artificial Intelligence
A.G.L.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 17:00
02.18.2026 17:00
Summary
Deep learning models for high-resolution remote sensing classification frequently face challenges related to data scarcity and severe class imbalance. This work proposes a unified StyleGAN3 architecture adapted for multispectral image synthesis and classification to address these limitations. The method integrates a dense classification head into the discriminator, transforming it into a dual-purpose architecture, and implements a class-balanced adversarial training strategy where the generator is conditioned to sample from a uniform class distribution. To evaluate the system, a domain-specific framework assesses FID, Precision, and Recall using a custom multispectral judge model. Experimental results on eight river basin datasets demonstrate that the proposed method achieves a mean Average Accuracy (AA) of 90.4\%, significantly outperforming standard baselines (ResNet, ViT) and previous GAN-based frameworks (ResBaGAN, EffBaGAN) in data-scarce scenarios.
Deep learning models for high-resolution remote sensing classification frequently face challenges related to data scarcity and severe class imbalance. This work proposes a unified StyleGAN3 architecture adapted for multispectral image synthesis and classification to address these limitations. The method integrates a dense classification head into the discriminator, transforming it into a dual-purpose architecture, and implements a class-balanced adversarial training strategy where the generator is conditioned to sample from a uniform class distribution. To evaluate the system, a domain-specific framework assesses FID, Precision, and Recall using a custom multispectral judge model. Experimental results on eight river basin datasets demonstrate that the proposed method achieves a mean Average Accuracy (AA) of 90.4\%, significantly outperforming standard baselines (ResNet, ViT) and previous GAN-based frameworks (ResBaGAN, EffBaGAN) in data-scarce scenarios.
Direction
Argüello Pedreira, Francisco Santiago (Tutorships)
Blanco Heras, Dora (Co-tutorships)
Argüello Pedreira, Francisco Santiago (Tutorships)
Blanco Heras, Dora (Co-tutorships)
Court
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
Design and implementation of an alert aggregation module in cybersecurity environments using machine learning
Authorship
R.G.B.
Master's Degree in Artificial Intelligence
R.G.B.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 10:00
02.18.2026 10:00
Summary
Security Operations Centers (SOCs) face increasing alert fatigue due to the high volume of alerts generated by modern security systems and the prevalence of false positives, which negatively impacts analysts’ efficiency and incident response. This Master’s Thesis addresses this challenge by proposing and evaluating a generic, configurable alert aggregation module developed within the SafeNet UEBA project, aimed at reducing alert volume while preserving structure and interpretability. The module integrates multiple aggregation strategies, including Self-Organizing Maps (SOM) and HDBSCAN clustering, optionally combined with UMAP for dimensionality reduction and visualization. It supports both single and hybrid aggregation modes and follows the Clean Architecture paradigm, ensuring modularity, extensibility and applicability beyond the SafeNet UEBA context. Experimental results on diverse cybersecurity datasets show that the proposed approach achieves high alert reduction rates while maintaining strong clustering quality. In addition to quantitative performance, the system provides analyst-oriented outputs such as visualizations, explainability summaries and metrics, supporting actionable insights in SOC environments where labeled data is unavailable.
Security Operations Centers (SOCs) face increasing alert fatigue due to the high volume of alerts generated by modern security systems and the prevalence of false positives, which negatively impacts analysts’ efficiency and incident response. This Master’s Thesis addresses this challenge by proposing and evaluating a generic, configurable alert aggregation module developed within the SafeNet UEBA project, aimed at reducing alert volume while preserving structure and interpretability. The module integrates multiple aggregation strategies, including Self-Organizing Maps (SOM) and HDBSCAN clustering, optionally combined with UMAP for dimensionality reduction and visualization. It supports both single and hybrid aggregation modes and follows the Clean Architecture paradigm, ensuring modularity, extensibility and applicability beyond the SafeNet UEBA context. Experimental results on diverse cybersecurity datasets show that the proposed approach achieves high alert reduction rates while maintaining strong clustering quality. In addition to quantitative performance, the system provides analyst-oriented outputs such as visualizations, explainability summaries and metrics, supporting actionable insights in SOC environments where labeled data is unavailable.
Direction
CARIÑENA AMIGO, MARIA PURIFICACION (Tutorships)
Piñón Blanco, Camilo (Co-tutorships)
CARIÑENA AMIGO, MARIA PURIFICACION (Tutorships)
Piñón Blanco, Camilo (Co-tutorships)
Court
TABOADA IGLESIAS, MARÍA JESÚS (Chairman)
VIDAL AGUIAR, JUAN CARLOS (Secretary)
Cotos Yáñez, José Manuel (Member)
TABOADA IGLESIAS, MARÍA JESÚS (Chairman)
VIDAL AGUIAR, JUAN CARLOS (Secretary)
Cotos Yáñez, José Manuel (Member)
Automated workflow for the exploration of CT based predictors of Atrial Fibrillation
Authorship
M.I.Z.
Master's Degree in Artificial Intelligence
M.I.Z.
Master's Degree in Artificial Intelligence
Defense date
02.19.2026 11:15
02.19.2026 11:15
Summary
Atrial fibrillation (AF) stands as the most prevalent cardiac arrhythmia, yet its early detection remains a significant clinical challenge due to its frequently asymptomatic nature. Once diagnosed, catheter ablation is considered a cornerstone therapy, however, recurrence rates remain high, driven by complex patient specific remodeling that makes predicting recurrence highly difficult. While various prognostic markers exist, clinical consensus is largely limited to global left atrial volume. Furthermore, extracting advanced markers from high resolution computed tomography (CT) is labor intensive and prone to inter observer variability. This work addresses these limitations by proposing a fully automated, end to end computational pipeline to quantify anatomical markers that the literature suggests are most significant for characterizing the AF substrate from CT scans. We employ a self configuring deep learning framework, trained on the public ImageCAS dataset, to segment the right atrium (RA), left atrium (LA), left atrial appendage (LAA), and pulmonary veins (PVs). A geometric post processing module automates the extraction of a predictor panel, global volumetry (LA, RA, and LAA volumes), shape estimation (LA Sphericity), regional remodeling indices (LA Posterior Volume, Posterior Anterior Volume Ratio (PAVR)), and PV topology (individual ostial perimeters and inter ostium distances). The pipeline was tested on a local cohort of 30 AF patients from the Hospital Clinico Universitario de Santiago (CHUS), achieving a segmentation exponential moving average Dice score of 0.9592 on the ImageCAS validation set. The automated analysis captured advanced remodeling profiles, including a mean LA volume 117.29 mL, a mean RA volume of 139.31 mL, and a mean LAA volume of 10.49 mL. Shape analysis identified a median LASP 85.08 percent, a median LA Posterior Volume of 36.31 mL, a median PAVR of 38.81 percent and identified anatomical variants such as Left Common Trunks in 33 percent of cases. This framework bridges the gap between raw imaging data and quantitative patient stratification, offering a scalable solution that can be extended to much larger cohorts to apply more advanced statistical analysis methods.
Atrial fibrillation (AF) stands as the most prevalent cardiac arrhythmia, yet its early detection remains a significant clinical challenge due to its frequently asymptomatic nature. Once diagnosed, catheter ablation is considered a cornerstone therapy, however, recurrence rates remain high, driven by complex patient specific remodeling that makes predicting recurrence highly difficult. While various prognostic markers exist, clinical consensus is largely limited to global left atrial volume. Furthermore, extracting advanced markers from high resolution computed tomography (CT) is labor intensive and prone to inter observer variability. This work addresses these limitations by proposing a fully automated, end to end computational pipeline to quantify anatomical markers that the literature suggests are most significant for characterizing the AF substrate from CT scans. We employ a self configuring deep learning framework, trained on the public ImageCAS dataset, to segment the right atrium (RA), left atrium (LA), left atrial appendage (LAA), and pulmonary veins (PVs). A geometric post processing module automates the extraction of a predictor panel, global volumetry (LA, RA, and LAA volumes), shape estimation (LA Sphericity), regional remodeling indices (LA Posterior Volume, Posterior Anterior Volume Ratio (PAVR)), and PV topology (individual ostial perimeters and inter ostium distances). The pipeline was tested on a local cohort of 30 AF patients from the Hospital Clinico Universitario de Santiago (CHUS), achieving a segmentation exponential moving average Dice score of 0.9592 on the ImageCAS validation set. The automated analysis captured advanced remodeling profiles, including a mean LA volume 117.29 mL, a mean RA volume of 139.31 mL, and a mean LAA volume of 10.49 mL. Shape analysis identified a median LASP 85.08 percent, a median LA Posterior Volume of 36.31 mL, a median PAVR of 38.81 percent and identified anatomical variants such as Left Common Trunks in 33 percent of cases. This framework bridges the gap between raw imaging data and quantitative patient stratification, offering a scalable solution that can be extended to much larger cohorts to apply more advanced statistical analysis methods.
Direction
NUÑEZ GARCIA, MARTA (Tutorships)
Rodríguez Mañero, Moises (Co-tutorships)
NUÑEZ GARCIA, MARTA (Tutorships)
Rodríguez Mañero, Moises (Co-tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
LAMA PENIN, MANUEL (Secretary)
VILA BLANCO, NICOLAS (Member)
GARCIA TAHOCES, PABLO (Chairman)
LAMA PENIN, MANUEL (Secretary)
VILA BLANCO, NICOLAS (Member)
A Comparative Analysis of LLM and Knowledge Graph Integration Strategies in the Medical Domain
Authorship
A.L.C.
Master's Degree in Artificial Intelligence
A.L.C.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 10:30
02.18.2026 10:30
Summary
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding, yet their application in high-stakes domains like clinical medicine is hindered by hallucinations and a lack of up-to-date domain-specific knowledge. Knowledge Graphs (KGs) offer structured, verifiable facts that can potentially ground these models, yet the optimal architecture for this integration remains an open research question. This work presents a comprehensive comparative analysis of three distinct integration strategies: (1) Retrieval-Augmented Generation (RAG) evaluating both lexical and semantic search, (2) Parameter-Efficient Fine-Tuning (LoRA) on ontology definitions, and (3) A novel Multi-Stage Reasoning Framework. Focused specifically on the cardiovascular domain, these approaches were evaluated using the Cardiovascular Disease Ontology (CVDO) and a specifically curated subset of the MedMCQA dataset, filtered by Named Entity Recognition (GLiNER) to ensure domain relevance with the context of the ontology. The results indicated mixed outcomes depending on the capacity of the model, where the proposed Retrieval-Augmented CoT (RA-CoT) strategy achieved global peak accuracy (93.23%) in a state-of-the-art model. This suggests that allowing models intermediate reasoning steps is critical and that combining multi-stage reasoning with external knowledge grounding can yield the strongest performance in our setting for complex clinical exam-style multiple-choice questions.
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding, yet their application in high-stakes domains like clinical medicine is hindered by hallucinations and a lack of up-to-date domain-specific knowledge. Knowledge Graphs (KGs) offer structured, verifiable facts that can potentially ground these models, yet the optimal architecture for this integration remains an open research question. This work presents a comprehensive comparative analysis of three distinct integration strategies: (1) Retrieval-Augmented Generation (RAG) evaluating both lexical and semantic search, (2) Parameter-Efficient Fine-Tuning (LoRA) on ontology definitions, and (3) A novel Multi-Stage Reasoning Framework. Focused specifically on the cardiovascular domain, these approaches were evaluated using the Cardiovascular Disease Ontology (CVDO) and a specifically curated subset of the MedMCQA dataset, filtered by Named Entity Recognition (GLiNER) to ensure domain relevance with the context of the ontology. The results indicated mixed outcomes depending on the capacity of the model, where the proposed Retrieval-Augmented CoT (RA-CoT) strategy achieved global peak accuracy (93.23%) in a state-of-the-art model. This suggests that allowing models intermediate reasoning steps is critical and that combining multi-stage reasoning with external knowledge grounding can yield the strongest performance in our setting for complex clinical exam-style multiple-choice questions.
Direction
CHAVES FRAGA, DAVID (Tutorships)
BUGARIN DIZ, ALBERTO JOSE (Co-tutorships)
LAMA PENIN, MANUEL (Co-tutorships)
CHAVES FRAGA, DAVID (Tutorships)
BUGARIN DIZ, ALBERTO JOSE (Co-tutorships)
LAMA PENIN, MANUEL (Co-tutorships)
Court
TABOADA IGLESIAS, MARÍA JESÚS (Chairman)
VIDAL AGUIAR, JUAN CARLOS (Secretary)
Cotos Yáñez, José Manuel (Member)
TABOADA IGLESIAS, MARÍA JESÚS (Chairman)
VIDAL AGUIAR, JUAN CARLOS (Secretary)
Cotos Yáñez, José Manuel (Member)
hEArT: A deep learning-based framework for automated quantification of Epicardial Adipose Tissue in multi-modal and multi-scale CT scans
Authorship
O.N.M.J.
Master's Degree in Artificial Intelligence
O.N.M.J.
Master's Degree in Artificial Intelligence
Defense date
02.19.2026 11:45
02.19.2026 11:45
Summary
Epicardial Adipose Tissue (EAT) is increasingly recognized as a significant factor in the pathophysiology of various cardiovascular diseases, suggesting that its accurate quantification could be of high clinical value. However, current reliance on manual segmentation from Computed Tomography (CT) scans presents challenges regarding efficiency and reproducibility, potentially limiting its broad application in research and clinical settings. This thesis presents the hEArT (heart Epicardial Adipose Tissue) framework, a novel, fully automated deep learning pipeline, aimed at addressing the complexities of EAT quantification across multi-modal and multi-scale CT scans. We propose a generalist to specialist architectural approach for pericardium segmentation, designed to delineate the anatomical boundaries necessary for isolating EAT. This process involves utilizing a large-scale foundation model to estimate the pericardium boundary to mitigate issues related to domain shift, followed by the application of a specialized corrector model designed to enhance anatomical precision in the final output. Additionally, as opposed to conventional EAT segmentation using image thresholding with fixed Hounsfield Unit values, our method estimates image-specific thresholds based on Gaussian Mixture Models. Comparative analysis suggests that this custom method may offer advantages over the state-of-the-art. Our framework was evaluated using a heterogeneous, real-world clinical dataset to assess its consistency and accuracy relative to semi-automated reference standards. Notably, in addition to being fully automated, the hEArt framework computed EAT volumes that showed stronger correlations with a panel of fat-related biological markers than those obtained using standard methods. These findings suggest that the proposed framework could serve as a scalable and objective tool, potentially facilitating more extensive clinical research into the biological relevance of EAT.
Epicardial Adipose Tissue (EAT) is increasingly recognized as a significant factor in the pathophysiology of various cardiovascular diseases, suggesting that its accurate quantification could be of high clinical value. However, current reliance on manual segmentation from Computed Tomography (CT) scans presents challenges regarding efficiency and reproducibility, potentially limiting its broad application in research and clinical settings. This thesis presents the hEArT (heart Epicardial Adipose Tissue) framework, a novel, fully automated deep learning pipeline, aimed at addressing the complexities of EAT quantification across multi-modal and multi-scale CT scans. We propose a generalist to specialist architectural approach for pericardium segmentation, designed to delineate the anatomical boundaries necessary for isolating EAT. This process involves utilizing a large-scale foundation model to estimate the pericardium boundary to mitigate issues related to domain shift, followed by the application of a specialized corrector model designed to enhance anatomical precision in the final output. Additionally, as opposed to conventional EAT segmentation using image thresholding with fixed Hounsfield Unit values, our method estimates image-specific thresholds based on Gaussian Mixture Models. Comparative analysis suggests that this custom method may offer advantages over the state-of-the-art. Our framework was evaluated using a heterogeneous, real-world clinical dataset to assess its consistency and accuracy relative to semi-automated reference standards. Notably, in addition to being fully automated, the hEArt framework computed EAT volumes that showed stronger correlations with a panel of fat-related biological markers than those obtained using standard methods. These findings suggest that the proposed framework could serve as a scalable and objective tool, potentially facilitating more extensive clinical research into the biological relevance of EAT.
Direction
NUÑEZ GARCIA, MARTA (Tutorships)
Eiras Penas, Sonia (Co-tutorships)
NUÑEZ GARCIA, MARTA (Tutorships)
Eiras Penas, Sonia (Co-tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
LAMA PENIN, MANUEL (Secretary)
VILA BLANCO, NICOLAS (Member)
GARCIA TAHOCES, PABLO (Chairman)
LAMA PENIN, MANUEL (Secretary)
VILA BLANCO, NICOLAS (Member)
Environmental assessment of biolipid production from sewage sludge by Yarrowia lipolytica through the application of Life Cycle Assessment within a circular economy context
Authorship
M.M.F.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
M.M.F.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
Defense date
02.19.2026 11:30
02.19.2026 11:30
Summary
The integrated assessment of environmental and economic impacts plays a fundamental role in the definition of strategies aligned with the principles of the circular economy, particularly in the search for biorefinery processes aimed at the production of bio-based products that enable a reduction in the use of fossil resources and support the transition towards more sustainable production systems. Within this framework, the present work provides an integrated evaluation of the technical, economic, and environmental viability of a sewage sludge valorisation process aimed at the production of biolipids through the cultivation of the oleaginous yeast Yarrowia lipolytica, within a biorefinery approach. The process has been scaled up from experimental laboratory data to an industrial-scale scenario through process modelling using SuperPro Designer, enabling the simulation of the different unit operations and the estimation of the main material and energy flows. Based on the developed model, a preliminary techno-economic analysis was carried out using classical economic indicators, such as net present value and payback period, in order to assess the economic feasibility of the process under the assumptions considered. In parallel, a cradle-to-gate Life Cycle Assessment was performed in accordance with ISO 14040 and ISO 14044 standards using the SimaPro software, with the aim of quantifying the environmental impacts associated with the overall process, as well as with the main product and the generated by-product, and of identifying the main critical points of the system. Additionally, a sensitivity analysis focused on the evaluation of alternative scenarios aimed at improving the environmental performance of the process was conducted. The study is further complemented by the calculation of environmental, circularity, and waste valorisation indicators, which allow the environmental performance of the biolipid and the by-product to be contextualized against their respective conventional alternatives. The results obtained demonstrate the potential of the analysed process from both a techno-economic and environmental perspective, showing a performance comparable to that of the considered conventional alternatives and positioning it as a strategy consistent with the principles of the circular economy.
The integrated assessment of environmental and economic impacts plays a fundamental role in the definition of strategies aligned with the principles of the circular economy, particularly in the search for biorefinery processes aimed at the production of bio-based products that enable a reduction in the use of fossil resources and support the transition towards more sustainable production systems. Within this framework, the present work provides an integrated evaluation of the technical, economic, and environmental viability of a sewage sludge valorisation process aimed at the production of biolipids through the cultivation of the oleaginous yeast Yarrowia lipolytica, within a biorefinery approach. The process has been scaled up from experimental laboratory data to an industrial-scale scenario through process modelling using SuperPro Designer, enabling the simulation of the different unit operations and the estimation of the main material and energy flows. Based on the developed model, a preliminary techno-economic analysis was carried out using classical economic indicators, such as net present value and payback period, in order to assess the economic feasibility of the process under the assumptions considered. In parallel, a cradle-to-gate Life Cycle Assessment was performed in accordance with ISO 14040 and ISO 14044 standards using the SimaPro software, with the aim of quantifying the environmental impacts associated with the overall process, as well as with the main product and the generated by-product, and of identifying the main critical points of the system. Additionally, a sensitivity analysis focused on the evaluation of alternative scenarios aimed at improving the environmental performance of the process was conducted. The study is further complemented by the calculation of environmental, circularity, and waste valorisation indicators, which allow the environmental performance of the biolipid and the by-product to be contextualized against their respective conventional alternatives. The results obtained demonstrate the potential of the analysed process from both a techno-economic and environmental perspective, showing a performance comparable to that of the considered conventional alternatives and positioning it as a strategy consistent with the principles of the circular economy.
Direction
MOREIRA VILAR, MARIA TERESA (Tutorships)
MOREIRA VILAR, MARIA TERESA (Tutorships)
Court
MOSQUERA CORRAL, ANUSKA (Chairman)
Pedrouso Fuentes, Alba (Secretary)
RODRIGUEZ MARTINEZ, HECTOR (Member)
MOSQUERA CORRAL, ANUSKA (Chairman)
Pedrouso Fuentes, Alba (Secretary)
RODRIGUEZ MARTINEZ, HECTOR (Member)
Pre-design of a heat recovery system for charcoal production kilns and comparison with biomass or natural gas heating systems
Authorship
J.M.R.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
J.M.R.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
Defense date
02.19.2026 12:00
02.19.2026 12:00
Summary
The production of charcoal has become a strategic alternative within the current context of decarbonization and the transition towards renewable raw materials. This resource, used as a reducing agent in the manufacture of ferroalloys such as ferrosilicon, offers environmental advantages compared to mineral coal. However, charcoal production involves energyintensive stages, among which the predrying of wood is particularly significant. This step is essential to ensure an adequate performance during pyrolysis, yet it represents a major thermal demand, typically covered by conventional solutions such as biomass or natural gas boilers, which entail relevant economic and environmental impacts. At the same time, pyrolysis furnaces generate hightemperature exhaust gases with considerable energy content that, in conventional plant configurations, are discharged directly to the atmosphere without any recovery. This presents a clear opportunity to integrate a heatrecovery system capable of reducing auxiliary fuel consumption and improving the overall energy efficiency of the facility. This Master’s Thesis analyses several alternatives for heating the water used in the wooddrying process, comparing three main options: biomass boiler, natural gas boiler, and heat recovery from furnace exhaust gases. To this end, energy balances, cost estimations, environmental assessments and a preliminary design of the selected solution are developed, including sizing of the equipment and material selection. The study concludes with a technical, economic and environmental comparison of all alternatives, ultimately identifying the most promising solution for integration into the future charcoal production plant.
The production of charcoal has become a strategic alternative within the current context of decarbonization and the transition towards renewable raw materials. This resource, used as a reducing agent in the manufacture of ferroalloys such as ferrosilicon, offers environmental advantages compared to mineral coal. However, charcoal production involves energyintensive stages, among which the predrying of wood is particularly significant. This step is essential to ensure an adequate performance during pyrolysis, yet it represents a major thermal demand, typically covered by conventional solutions such as biomass or natural gas boilers, which entail relevant economic and environmental impacts. At the same time, pyrolysis furnaces generate hightemperature exhaust gases with considerable energy content that, in conventional plant configurations, are discharged directly to the atmosphere without any recovery. This presents a clear opportunity to integrate a heatrecovery system capable of reducing auxiliary fuel consumption and improving the overall energy efficiency of the facility. This Master’s Thesis analyses several alternatives for heating the water used in the wooddrying process, comparing three main options: biomass boiler, natural gas boiler, and heat recovery from furnace exhaust gases. To this end, energy balances, cost estimations, environmental assessments and a preliminary design of the selected solution are developed, including sizing of the equipment and material selection. The study concludes with a technical, economic and environmental comparison of all alternatives, ultimately identifying the most promising solution for integration into the future charcoal production plant.
Direction
RODIL RODRIGUEZ, EVA (Tutorships)
Vázquez Varela, Alfonso (Co-tutorships)
RODIL RODRIGUEZ, EVA (Tutorships)
Vázquez Varela, Alfonso (Co-tutorships)
Court
MOSQUERA CORRAL, ANUSKA (Chairman)
Pedrouso Fuentes, Alba (Secretary)
RODRIGUEZ MARTINEZ, HECTOR (Member)
MOSQUERA CORRAL, ANUSKA (Chairman)
Pedrouso Fuentes, Alba (Secretary)
RODRIGUEZ MARTINEZ, HECTOR (Member)
VISTA (Visual Intelligence for Scene-Based Training and Assessment)
Authorship
M.O.N.
Master's Degree in Artificial Intelligence
M.O.N.
Master's Degree in Artificial Intelligence
Defense date
02.19.2026 12:15
02.19.2026 12:15
Summary
As the global population ages, the demand for scalable, non-pharmacological interventions for cognitive decline has intensified. This Master's Thesis presents VISTA (Visual Intelligence for Scene-Based Training and Assessment), a novel framework that integrates neuro-symbolic visual reasoning with Reinforcement Learning (RL) to deliver adaptive cognitive training. This Master's Thesis proposes a dual-engine architecture: a perception engine that utilizes a modified Multimodal Context-aware Consistency Alignment network to generate factually grounded exercises from visual scenes, and a tutoring engine driven by a Deep Q-Network (DQN) that dynamically adjusts curriculum difficulty. Quantitative evaluation on a custom CLEVR-based dataset reveals that the cosine similarity attention mechanism achieves 95.15\% accuracy, significantly outperforming standard baselines in complex logical reasoning tasks. Furthermore, the RL tutoring agent demonstrates robust adaptability, effectively tailoring learning trajectories for struggling, standard, and gifted user profiles without explicit rule-based programming. Experimentation and validation of the system was conducted in a simulated environment, allowing for reproducibility in the obtained results.
As the global population ages, the demand for scalable, non-pharmacological interventions for cognitive decline has intensified. This Master's Thesis presents VISTA (Visual Intelligence for Scene-Based Training and Assessment), a novel framework that integrates neuro-symbolic visual reasoning with Reinforcement Learning (RL) to deliver adaptive cognitive training. This Master's Thesis proposes a dual-engine architecture: a perception engine that utilizes a modified Multimodal Context-aware Consistency Alignment network to generate factually grounded exercises from visual scenes, and a tutoring engine driven by a Deep Q-Network (DQN) that dynamically adjusts curriculum difficulty. Quantitative evaluation on a custom CLEVR-based dataset reveals that the cosine similarity attention mechanism achieves 95.15\% accuracy, significantly outperforming standard baselines in complex logical reasoning tasks. Furthermore, the RL tutoring agent demonstrates robust adaptability, effectively tailoring learning trajectories for struggling, standard, and gifted user profiles without explicit rule-based programming. Experimentation and validation of the system was conducted in a simulated environment, allowing for reproducibility in the obtained results.
Direction
CATALA BOLOS, ALEJANDRO (Tutorships)
CONDORI FERNANDEZ, OLINDA NELLY (Co-tutorships)
CATALA BOLOS, ALEJANDRO (Tutorships)
CONDORI FERNANDEZ, OLINDA NELLY (Co-tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
LAMA PENIN, MANUEL (Secretary)
VILA BLANCO, NICOLAS (Member)
GARCIA TAHOCES, PABLO (Chairman)
LAMA PENIN, MANUEL (Secretary)
VILA BLANCO, NICOLAS (Member)
Strategies for managing greenhouse gas emissions in the border region of San Diego, California and Tijuana, Baja California. Impact of urban mobility and the integration of clean technologies
Authorship
E.P.A.
Master's Degree in Environmental Engineering (3rd ed)
E.P.A.
Master's Degree in Environmental Engineering (3rd ed)
Defense date
02.23.2026 13:00
02.23.2026 13:00
Summary
Tijuana, Mexico, and San Diego, USA, two countries, two states, one region. Often mistakenly called twin cities, they share not only history, culture, and tradition but also significant economic, political, and environmental challenges, one of which is the main focus of this research: air quality, public health, and their correlation with greenhouse gases (GHG). Both cities, as centers of high industrial activity (maquiladora industry) and vehicular traffic, are responsible for a significant proportion of these emissions. While their daily interaction compels hundreds of thousands of people to cross the border daily for various reasons and by different means, comparing the effects of these gases in both cities is unavoidable. Studying greenhouse gas emissions in two cities as close as Tijuana and San Diego allows us to understand how urban contexts influence their environmental impact. These two cities undoubtedly provide an ideal framework for analyzing factors such as population growth, mobility, industrial activity, and public and cooperative policies and their influence on the generation of polluting emissions. This study aims to identify the main sources of greenhouse gas emissions in the region, as well as evaluate and compare the different mitigation strategies implemented and identify best practices that can be replicated and adapted between the two cities, thus strengthening bilateral cooperation.
Tijuana, Mexico, and San Diego, USA, two countries, two states, one region. Often mistakenly called twin cities, they share not only history, culture, and tradition but also significant economic, political, and environmental challenges, one of which is the main focus of this research: air quality, public health, and their correlation with greenhouse gases (GHG). Both cities, as centers of high industrial activity (maquiladora industry) and vehicular traffic, are responsible for a significant proportion of these emissions. While their daily interaction compels hundreds of thousands of people to cross the border daily for various reasons and by different means, comparing the effects of these gases in both cities is unavoidable. Studying greenhouse gas emissions in two cities as close as Tijuana and San Diego allows us to understand how urban contexts influence their environmental impact. These two cities undoubtedly provide an ideal framework for analyzing factors such as population growth, mobility, industrial activity, and public and cooperative policies and their influence on the generation of polluting emissions. This study aims to identify the main sources of greenhouse gas emissions in the region, as well as evaluate and compare the different mitigation strategies implemented and identify best practices that can be replicated and adapted between the two cities, thus strengthening bilateral cooperation.
Direction
HOSPIDO QUINTANA, ALMUDENA (Tutorships)
HOSPIDO QUINTANA, ALMUDENA (Tutorships)
Court
Omil Prieto, Francisco (Chairman)
OTERO PEREZ, XOSE LOIS (Secretary)
GONZALEZ GARCIA, SARA (Member)
Omil Prieto, Francisco (Chairman)
OTERO PEREZ, XOSE LOIS (Secretary)
GONZALEZ GARCIA, SARA (Member)
Protobiome: Prototypical Networks for Few-Shot Multiclass Classification of the Oral Microbiome
Authorship
E.P.V.
Master's Degree in Artificial Intelligence
E.P.V.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 11:00
02.18.2026 11:00
Summary
Oral microbiome datasets are characterized by high dimensionality, sparsity, and compositionality, which often lead to poor generalization in traditional machine learning models under low-sample regimes. In this work, we present Protobiome, a framework utilizing Prototypical Networks for few-shot multiclass classification across saliva, subgingival, and supragingival niches. Our approach integrates specialized preprocessing, feature selection, and data augmentation tailored for the simplex space. Results show that Prototypical Networks significantly outperform state-of-the-art models like XGBoost in minority class detection. For saliva samples, our model achieved a Gingivitis F1-score of 0.958 with perfect precision and specificity using less than 20 samples. Explainability analysis via UMAP and SHAP values confirms that the model identifies a coherent microbial hierarchy, with specific Amplicon Sequence Variants (ASVs) serving as discriminative indicators of the actual condition. This study demonstrates that metric-based few-shot learning provides a robust, interpretable foundation for non-invasive clinical diagnostics.
Oral microbiome datasets are characterized by high dimensionality, sparsity, and compositionality, which often lead to poor generalization in traditional machine learning models under low-sample regimes. In this work, we present Protobiome, a framework utilizing Prototypical Networks for few-shot multiclass classification across saliva, subgingival, and supragingival niches. Our approach integrates specialized preprocessing, feature selection, and data augmentation tailored for the simplex space. Results show that Prototypical Networks significantly outperform state-of-the-art models like XGBoost in minority class detection. For saliva samples, our model achieved a Gingivitis F1-score of 0.958 with perfect precision and specificity using less than 20 samples. Explainability analysis via UMAP and SHAP values confirms that the model identifies a coherent microbial hierarchy, with specific Amplicon Sequence Variants (ASVs) serving as discriminative indicators of the actual condition. This study demonstrates that metric-based few-shot learning provides a robust, interpretable foundation for non-invasive clinical diagnostics.
Direction
VILA BLANCO, NICOLAS (Tutorships)
TOMAS CARMONA, INMACULADA (Co-tutorships)
LAMAS PEREZ, JOSE MANUEL (Co-tutorships)
VILA BLANCO, NICOLAS (Tutorships)
TOMAS CARMONA, INMACULADA (Co-tutorships)
LAMAS PEREZ, JOSE MANUEL (Co-tutorships)
Court
TABOADA IGLESIAS, MARÍA JESÚS (Chairman)
VIDAL AGUIAR, JUAN CARLOS (Secretary)
Cotos Yáñez, José Manuel (Member)
TABOADA IGLESIAS, MARÍA JESÚS (Chairman)
VIDAL AGUIAR, JUAN CARLOS (Secretary)
Cotos Yáñez, José Manuel (Member)
Monocular 6D Object Pose Estimation for Mixed Reality Applications
Authorship
D.P.
Master's Degree in Computer Vision
D.P.
Master's Degree in Computer Vision
Defense date
02.04.2026 09:50
02.04.2026 09:50
Summary
This masters thesis report presents a complete pipeline for the real-time 6D object pose estimation on a standalone mixed reality hardware, specifically the Meta Quest 3 headset. Unlike traditional 2D object detection, 6D pose estimation recovers both 3D position and 3D orientations of an object. It enables presice spatial understanding that is crucial for mixed reality applications, such as assembly guidance, accessibility and entertainment. This work leverages the recently released Meta's Passthrough Camera API to implement computer vision tasks directly on the device. The proposed system consists of three main components: (1) a procedural synthetic data generation pipeline utilizing Python and Blender to create photorealistic training images with pixel-perfect 6D annotations; (2) an implementation of the lightweight YOLOX-6D-Pose architecture optimized for edge inference; and (3) a Unity-based mixed reality application using the Unity Sentis inference engine. Experimental results demonstrate a successful Sim-to-Real transfer, achieving a BOP Mean Average Recall (AR) of 62.79% on real-world data without using any real training images. Ablation study confirms that domain randomization is important and improves performance by over 12%. Furthermore, dynamic INT8 quantization reduced the model size by around 75% and inference latency to 201ms with very little accuracy loss. This work validates the possibility of performing 6D pose estimation on consumer VR headsets, opening the way for spatially aware MR applications in many different applications.
This masters thesis report presents a complete pipeline for the real-time 6D object pose estimation on a standalone mixed reality hardware, specifically the Meta Quest 3 headset. Unlike traditional 2D object detection, 6D pose estimation recovers both 3D position and 3D orientations of an object. It enables presice spatial understanding that is crucial for mixed reality applications, such as assembly guidance, accessibility and entertainment. This work leverages the recently released Meta's Passthrough Camera API to implement computer vision tasks directly on the device. The proposed system consists of three main components: (1) a procedural synthetic data generation pipeline utilizing Python and Blender to create photorealistic training images with pixel-perfect 6D annotations; (2) an implementation of the lightweight YOLOX-6D-Pose architecture optimized for edge inference; and (3) a Unity-based mixed reality application using the Unity Sentis inference engine. Experimental results demonstrate a successful Sim-to-Real transfer, achieving a BOP Mean Average Recall (AR) of 62.79% on real-world data without using any real training images. Ablation study confirms that domain randomization is important and improves performance by over 12%. Furthermore, dynamic INT8 quantization reduced the model size by around 75% and inference latency to 201ms with very little accuracy loss. This work validates the possibility of performing 6D pose estimation on consumer VR headsets, opening the way for spatially aware MR applications in many different applications.
Direction
FLORES GONZALEZ, JULIAN CARLOS (Tutorships)
Glowacki , David Ryan (Co-tutorships)
FLORES GONZALEZ, JULIAN CARLOS (Tutorships)
Glowacki , David Ryan (Co-tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
Design and validation of a trustworthy conversational assistant supported by retrieval-augmented generation techniques
Authorship
D.R.M.
Master's Degree in Artificial Intelligence
D.R.M.
Master's Degree in Artificial Intelligence
Defense date
02.03.2026 13:00
02.03.2026 13:00
Summary
Large Language Models have shown remarkable capabilities in natural language understanding and generation, however, their utilization in sensitive domains such as medicine or law raises concerns regarding reliability, transparency, and trustworthiness. In response to these challenges, Retrieval-Augmented Generation (RAG) techniques have emerged to ground generated outputs in reliable documentary source data. This Master's thesis proposes a RAG architecture which extends the base pipeline by integrating multiple safeguard mechanisms, including a query verification module to detect out-of-scope questions, a multi-source retrieval system along with knowledge-refinement mechanisms to optimize information extraction from the knowledge base and a chain-of-thought module to enhance the reasoning capabilities of the generator model. Across the three experiments conducted, the results demonstrate that the inclusion of each safeguard mechanism improves system performance under different settings. Specifically, the proposed architecture more accurately rejects out-of-scope questions, enhances retrieval effectiveness compared to widely used retrievers such as BM25. It also achieves improved performance on multiple-choice question answering and fact-checking tasks through its dual vector-store design and chain-of-thought module.
Large Language Models have shown remarkable capabilities in natural language understanding and generation, however, their utilization in sensitive domains such as medicine or law raises concerns regarding reliability, transparency, and trustworthiness. In response to these challenges, Retrieval-Augmented Generation (RAG) techniques have emerged to ground generated outputs in reliable documentary source data. This Master's thesis proposes a RAG architecture which extends the base pipeline by integrating multiple safeguard mechanisms, including a query verification module to detect out-of-scope questions, a multi-source retrieval system along with knowledge-refinement mechanisms to optimize information extraction from the knowledge base and a chain-of-thought module to enhance the reasoning capabilities of the generator model. Across the three experiments conducted, the results demonstrate that the inclusion of each safeguard mechanism improves system performance under different settings. Specifically, the proposed architecture more accurately rejects out-of-scope questions, enhances retrieval effectiveness compared to widely used retrievers such as BM25. It also achieves improved performance on multiple-choice question answering and fact-checking tasks through its dual vector-store design and chain-of-thought module.
Direction
ALONSO MORAL, JOSE MARIA (Tutorships)
CATALA BOLOS, ALEJANDRO (Co-tutorships)
ALONSO MORAL, JOSE MARIA (Tutorships)
CATALA BOLOS, ALEJANDRO (Co-tutorships)
Court
BUGARIN DIZ, ALBERTO JOSE (Chairman)
VALLADARES RODRIGUEZ, SONIA MARIA (Secretary)
PICHEL CAMPOS, JOSE RAMON (Member)
BUGARIN DIZ, ALBERTO JOSE (Chairman)
VALLADARES RODRIGUEZ, SONIA MARIA (Secretary)
PICHEL CAMPOS, JOSE RAMON (Member)
Decarbonization of energy intensive processes through life cycle assessment: superheated steam wood dryer
Authorship
S.R.P.
Master's Degree in Environmental Engineering (3rd ed)
S.R.P.
Master's Degree in Environmental Engineering (3rd ed)
Defense date
02.23.2026 12:30
02.23.2026 12:30
Summary
The wood drying phase is characterized by a high energy consumption, which represents around 60% of the total consumption of the industry. Kiln drying is the dominant technology used to reduce the moisture content in the planks and reach industrial standards. Currently, most of the kilns make use of combustion processes for heat generation, hence why developing more efficient and electrified alternatives is key to reduce the direct emissions of the industry. The thesis was focused around the design of a kiln dryer which, instead of hot air, uses superheated steam as the drying agent. This emerging technology can notably reduce the carbon emissions thanks to two mechanisms: (1) recovery of the latent heat of the vaporized water from the planks, through a pressure increase and a heat exchanger, and (2) the possibility of electrifying the generation and heat recovery through technologies such as Mechanical Vapor Recompression or heat pumps. A mathematical model was developed using Python to simulate the heat and mass transfer in the drying process, evaluating operation conditions and the moisture content evolution in the wood stack. Simultaneously, the heat recovery system was simulated using Aspen PLUS. Finally, the reduction in carbon footprint and the environmental profile were evaluated through Life Cycle Assessment, comparing the data with the conventional process making use of bibliographical and/or simulated data.
The wood drying phase is characterized by a high energy consumption, which represents around 60% of the total consumption of the industry. Kiln drying is the dominant technology used to reduce the moisture content in the planks and reach industrial standards. Currently, most of the kilns make use of combustion processes for heat generation, hence why developing more efficient and electrified alternatives is key to reduce the direct emissions of the industry. The thesis was focused around the design of a kiln dryer which, instead of hot air, uses superheated steam as the drying agent. This emerging technology can notably reduce the carbon emissions thanks to two mechanisms: (1) recovery of the latent heat of the vaporized water from the planks, through a pressure increase and a heat exchanger, and (2) the possibility of electrifying the generation and heat recovery through technologies such as Mechanical Vapor Recompression or heat pumps. A mathematical model was developed using Python to simulate the heat and mass transfer in the drying process, evaluating operation conditions and the moisture content evolution in the wood stack. Simultaneously, the heat recovery system was simulated using Aspen PLUS. Finally, the reduction in carbon footprint and the environmental profile were evaluated through Life Cycle Assessment, comparing the data with the conventional process making use of bibliographical and/or simulated data.
Direction
MOREIRA VILAR, MARIA TERESA (Tutorships)
FEIJOO COSTA, GUMERSINDO (Co-tutorships)
MOREIRA VILAR, MARIA TERESA (Tutorships)
FEIJOO COSTA, GUMERSINDO (Co-tutorships)
Court
Omil Prieto, Francisco (Chairman)
OTERO PEREZ, XOSE LOIS (Secretary)
GONZALEZ GARCIA, SARA (Member)
Omil Prieto, Francisco (Chairman)
OTERO PEREZ, XOSE LOIS (Secretary)
GONZALEZ GARCIA, SARA (Member)
Towards sustainable fisheries management in the Atlantic and Mediterranean seas: an ecosystem services assessment framework
Authorship
R.R.S.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
R.R.S.
Master's Degree in Chemical Engineering and Bioprocess (2nd ed)
Defense date
02.19.2026 12:30
02.19.2026 12:30
Summary
The sustainability of capture fisheries represents a priority challenge in a context of increasing pressures on marine ecosystems and fishery resources. Addressing this challenge requires integrated assessment approaches capable of characterising not only the environmental performance of fisheries, but also the socio-economic and governance dimensions that condition their long-term viability. Within this framework, this study develops an integrated indicator framework aimed at assessing the sustainability of fishing activities from an ecosystem-based perspective. The proposed framework is built upon the integration of scientific and technical-institutional reference sources related to fishing activities and is articulated through a hierarchical structure that enables the integration of information across different levels of analysis. This approach facilitates a coherent and integrated interpretation of results, enhances their interpretability, and improves the usefulness of the framework as a decision-support tool for fisheries assessment and management processes. The applicability of the framework is demonstrated through its implementation in two European marine regions with contrasting ecological and management contexts: the Bay of Biscay and the Iberian Coast, and the Western Mediterranean. The analysis combines an integrated regional-scale assessment of fisheries with a species-level evaluation for three representative populations. The results reveal significant regional differences in terms of sustainability. The Bay of Biscay and the Iberian Coast generally exhibit a more favourable resource status, associated with higher stock productivity and stronger socio-economic performance. In contrast, the Western Mediterranean shows a less favourable environmental performance, characterised by more intense pressures on resources and particularly critical situations for certain populations, such as European hake. Overall, the results demonstrate the capacity of the indicator framework to identify critical sustainability issues, compare performance across regions, and provide evidence to support the design and adjustment of fisheries policies based on scientific criteria. The integrated approach adopted offers a coherent analytical basis for sector monitoring and for guiding adaptive management strategies aimed at strengthening the resilience of marine ecosystems and fisheries systems.
The sustainability of capture fisheries represents a priority challenge in a context of increasing pressures on marine ecosystems and fishery resources. Addressing this challenge requires integrated assessment approaches capable of characterising not only the environmental performance of fisheries, but also the socio-economic and governance dimensions that condition their long-term viability. Within this framework, this study develops an integrated indicator framework aimed at assessing the sustainability of fishing activities from an ecosystem-based perspective. The proposed framework is built upon the integration of scientific and technical-institutional reference sources related to fishing activities and is articulated through a hierarchical structure that enables the integration of information across different levels of analysis. This approach facilitates a coherent and integrated interpretation of results, enhances their interpretability, and improves the usefulness of the framework as a decision-support tool for fisheries assessment and management processes. The applicability of the framework is demonstrated through its implementation in two European marine regions with contrasting ecological and management contexts: the Bay of Biscay and the Iberian Coast, and the Western Mediterranean. The analysis combines an integrated regional-scale assessment of fisheries with a species-level evaluation for three representative populations. The results reveal significant regional differences in terms of sustainability. The Bay of Biscay and the Iberian Coast generally exhibit a more favourable resource status, associated with higher stock productivity and stronger socio-economic performance. In contrast, the Western Mediterranean shows a less favourable environmental performance, characterised by more intense pressures on resources and particularly critical situations for certain populations, such as European hake. Overall, the results demonstrate the capacity of the indicator framework to identify critical sustainability issues, compare performance across regions, and provide evidence to support the design and adjustment of fisheries policies based on scientific criteria. The integrated approach adopted offers a coherent analytical basis for sector monitoring and for guiding adaptive management strategies aimed at strengthening the resilience of marine ecosystems and fisheries systems.
Direction
MOREIRA VILAR, MARIA TERESA (Tutorships)
MOREIRA VILAR, MARIA TERESA (Tutorships)
Court
MOSQUERA CORRAL, ANUSKA (Chairman)
Pedrouso Fuentes, Alba (Secretary)
RODRIGUEZ MARTINEZ, HECTOR (Member)
MOSQUERA CORRAL, ANUSKA (Chairman)
Pedrouso Fuentes, Alba (Secretary)
RODRIGUEZ MARTINEZ, HECTOR (Member)
Evaluation of Neural Network Acceleration Tools and Optimization Techniques for FPGA-Based Edge Deployments
Authorship
Y.S.I.
Master's Degree in Artificial Intelligence
Y.S.I.
Master's Degree in Artificial Intelligence
Defense date
02.18.2026 17:30
02.18.2026 17:30
Summary
Industrial demand for deploying efficient neural network models on FPGA-based edge systems motivates this work, where throughput and energy efficiency are critical constraints. To address this challenge, a unified optimization pipeline integrating one-step pruning, retraining, post-training INT8 quantization, and FPGA compilation is developed and evaluated using AMD Vitis AI. The pipeline is assessed on three representative models spanning computer vision and natural language processing tasks. ResNet-18 and ENet are used to evaluate the impact of pruning and quantization on vision-oriented models, while BERT-Large is included to assess the applicability of the workflow beyond convolutional architectures. Experimental results show that structured pruning significantly improves FPGA throughput and energy efficiency for vision models, with ResNet-18 increases on-board throughput from 418.02 to 637.98 FPS (+52.6%) and improves energy efficiency from 12.3 to 20.4 FPS/W. For ENet, pruning and INT8 quantization improve host-side throughput from 27.61 to 34.25 FPS (+24.0%) with a 1.10-point mIoU reduction (93.43% to 92.33%), while on-board dual-thread execution reaches up to 17.22 FPS and 0.65 FPS/W versus 12.81 FPS and 0.44 FPS/W for the baseline. In contrast, while INT8 quantization preserves accuracy for BERT-Large on the host platform, FPGA compilation cannot be completed due to limited compiler support for transformer-specific operators. These results highlight that host-side evaluation alone may be insufficient to characterize deployment efficiency, motivating hardware-level measurements. While pruning and quantization effectively improve on-board performance for vision models, transformer deployment on FPGAs remains constrained by current compilation and operator support.
Industrial demand for deploying efficient neural network models on FPGA-based edge systems motivates this work, where throughput and energy efficiency are critical constraints. To address this challenge, a unified optimization pipeline integrating one-step pruning, retraining, post-training INT8 quantization, and FPGA compilation is developed and evaluated using AMD Vitis AI. The pipeline is assessed on three representative models spanning computer vision and natural language processing tasks. ResNet-18 and ENet are used to evaluate the impact of pruning and quantization on vision-oriented models, while BERT-Large is included to assess the applicability of the workflow beyond convolutional architectures. Experimental results show that structured pruning significantly improves FPGA throughput and energy efficiency for vision models, with ResNet-18 increases on-board throughput from 418.02 to 637.98 FPS (+52.6%) and improves energy efficiency from 12.3 to 20.4 FPS/W. For ENet, pruning and INT8 quantization improve host-side throughput from 27.61 to 34.25 FPS (+24.0%) with a 1.10-point mIoU reduction (93.43% to 92.33%), while on-board dual-thread execution reaches up to 17.22 FPS and 0.65 FPS/W versus 12.81 FPS and 0.44 FPS/W for the baseline. In contrast, while INT8 quantization preserves accuracy for BERT-Large on the host platform, FPGA compilation cannot be completed due to limited compiler support for transformer-specific operators. These results highlight that host-side evaluation alone may be insufficient to characterize deployment efficiency, motivating hardware-level measurements. While pruning and quantization effectively improve on-board performance for vision models, transformer deployment on FPGAs remains constrained by current compilation and operator support.
Direction
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
Losada Sanisidro, Pablo (Co-tutorships)
Febles Rodríguez, Adriana (Co-tutorships)
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
Losada Sanisidro, Pablo (Co-tutorships)
Febles Rodríguez, Adriana (Co-tutorships)
Court
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
IGLESIAS RODRIGUEZ, ROBERTO (Chairman)
SANTOS MATEOS, ROI (Secretary)
CARIÑENA AMIGO, MARIA PURIFICACION (Member)
Scanpath Prediction from Implicit Cues in Noisy Gaze Data
Authorship
L.U.F.
Master's Degree in Computer Vision
L.U.F.
Master's Degree in Computer Vision
Defense date
02.04.2026 09:30
02.04.2026 09:30
Summary
Eye-tracking is a vital tool for psychological and psychophysiological research; however, obtaining reliable data typically requires expensive equipment and controlled laboratory environments. While more affordable alternatives have been developed, they often lack the precision and sampling rates necessary for rigorous scientific study. In this work, we propose a model that integrates noisy, low-sampling-rate eye-tracking data with stimulus image features to reconstruct sequences of fixation centroids and their corresponding durations. Our approach aims to produce data that maintains the statistical properties of high-end tracking systems. To achieve this, we utilized the CocoFreeView dataset to generate realistic eye-tracking samples and developed a noise model that simulates the characteristics of widely used commercial eye-trackers. Finally, we leverage a Transformer-based architecture featuring a DINOv3 image encoder to recover the original fixation information.
Eye-tracking is a vital tool for psychological and psychophysiological research; however, obtaining reliable data typically requires expensive equipment and controlled laboratory environments. While more affordable alternatives have been developed, they often lack the precision and sampling rates necessary for rigorous scientific study. In this work, we propose a model that integrates noisy, low-sampling-rate eye-tracking data with stimulus image features to reconstruct sequences of fixation centroids and their corresponding durations. Our approach aims to produce data that maintains the statistical properties of high-end tracking systems. To achieve this, we utilized the CocoFreeView dataset to generate realistic eye-tracking samples and developed a noise model that simulates the characteristics of widely used commercial eye-trackers. Finally, we leverage a Transformer-based architecture featuring a DINOv3 image encoder to recover the original fixation information.
Direction
CORES COSTA, DANIEL (Tutorships)
CORES COSTA, DANIEL (Tutorships)
Court
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)
GARCIA TAHOCES, PABLO (Chairman)
BREA SANCHEZ, VICTOR MANUEL (Secretary)
López Martínez, Paula (Member)