Avoiding catastrophic forgetting in a continual learning framework for tabular data regression problems
Authorship
M.A.G.
Double Bachelor's Degree in Informatics Engineering and Mathematics
M.A.G.
Double Bachelor's Degree in Informatics Engineering and Mathematics
Defense date
02.20.2026 09:45
02.20.2026 09:45
Summary
This Bachelor's Thesis presents a continual and incremental learning solution for multivariate regression problems with tabular data. The research focuses on the adaptation and extension of the TRIL3 framework, whose original functionality was limited to classification, to address regression scenarios in a task-free context. The proposed methodology combines the use of the XuILVQ prototype model for the generation of synthetic data with mixture density networks as a predictive model. This approach allows for the mitigation of catastrophic forgetting in online learning environments and with the presence of concept drift. The effectiveness of the system was validated through a battery of experiments with reference datasets, contrasting the results with the state of the art. The results obtained demonstrate that the adapted solution maintains high robustness against forgetting and superior memory efficiency, with a very reduced ratio of stored prototypes relative to the total volume of data. Due to its characteristics, the proposal is especially suitable for its implementation in edge computing devices within the Industry 4.0 paradigm.
This Bachelor's Thesis presents a continual and incremental learning solution for multivariate regression problems with tabular data. The research focuses on the adaptation and extension of the TRIL3 framework, whose original functionality was limited to classification, to address regression scenarios in a task-free context. The proposed methodology combines the use of the XuILVQ prototype model for the generation of synthetic data with mixture density networks as a predictive model. This approach allows for the mitigation of catastrophic forgetting in online learning environments and with the presence of concept drift. The effectiveness of the system was validated through a battery of experiments with reference datasets, contrasting the results with the state of the art. The results obtained demonstrate that the adapted solution maintains high robustness against forgetting and superior memory efficiency, with a very reduced ratio of stored prototypes relative to the total volume of data. Due to its characteristics, the proposal is especially suitable for its implementation in edge computing devices within the Industry 4.0 paradigm.
Direction
MERA PEREZ, DAVID (Tutorships)
Fernández Castro, Bruno (Co-tutorships)
MERA PEREZ, DAVID (Tutorships)
Fernández Castro, Bruno (Co-tutorships)
Court
Barro Ameneiro, Senén (Chairman)
GARCIA FERNANDEZ, JULIAN (Secretary)
LADRA GONZALEZ, MANUEL EULOGIO (Member)
Barro Ameneiro, Senén (Chairman)
GARCIA FERNANDEZ, JULIAN (Secretary)
LADRA GONZALEZ, MANUEL EULOGIO (Member)
Stochastic Gradient Methods
Authorship
M.A.G.
Double Bachelor's Degree in Informatics Engineering and Mathematics
M.A.G.
Double Bachelor's Degree in Informatics Engineering and Mathematics
Defense date
02.12.2026 09:00
02.12.2026 09:00
Summary
This undergraduate thesis examines Stochastic Gradient Descent (SGD) methods as a cornerstone for solving large-scale optimization problems, particularly within the field of machine learning. Traditional gradient descent (GD) faces prohibitive computational limitations when dealing with massive datasets, as it requires processing the entire set of observations at each iteration. As an alternative, SGD employs noisy but efficient gradient estimates derived from random samples. Throughout this work, the regularity conditions necessary to ensure the method's stability, such as smoothness in expectation and variance control, are studied. Convergence results are analyzed for convex and strongly convex functions, as well as those satisfying the Polyak-Lojasiewicz condition. Finally, highly practical variants such as minibatch SGD and the momentum method are explored, justifying how these strategies mitigate noise and enhance optimization dynamics in ill-conditioned or over-parameterized scenarios.
This undergraduate thesis examines Stochastic Gradient Descent (SGD) methods as a cornerstone for solving large-scale optimization problems, particularly within the field of machine learning. Traditional gradient descent (GD) faces prohibitive computational limitations when dealing with massive datasets, as it requires processing the entire set of observations at each iteration. As an alternative, SGD employs noisy but efficient gradient estimates derived from random samples. Throughout this work, the regularity conditions necessary to ensure the method's stability, such as smoothness in expectation and variance control, are studied. Convergence results are analyzed for convex and strongly convex functions, as well as those satisfying the Polyak-Lojasiewicz condition. Finally, highly practical variants such as minibatch SGD and the momentum method are explored, justifying how these strategies mitigate noise and enhance optimization dynamics in ill-conditioned or over-parameterized scenarios.
Direction
GONZALEZ DIAZ, JULIO (Tutorships)
GONZALEZ DIAZ, JULIO (Tutorships)
Court
VAZQUEZ CENDON, MARIA ELENA (Chairman)
LOPEZ SOMOZA, LUCIA (Secretary)
CASAS MENDEZ, BALBINA VIRGINIA (Member)
VAZQUEZ CENDON, MARIA ELENA (Chairman)
LOPEZ SOMOZA, LUCIA (Secretary)
CASAS MENDEZ, BALBINA VIRGINIA (Member)
Gaming Platform for the Development of Different Competencies in Children
Authorship
S.C.B.B.
Bachelor’s Degree in Informatics Engineering
S.C.B.B.
Bachelor’s Degree in Informatics Engineering
Defense date
02.18.2026 09:30
02.18.2026 09:30
Summary
This Bachelor’s Final Degree Project, developed in collaboration with Pictoaplicaciones, aims to design and implement a multiplatform video game system focused on providing accessible learning and entertainment experiences, using clear and easily understandable visual language for children. Special attention is given to the inclusion of users with functional diversity and neurodivergent profiles, such as Autism Spectrum Disorder (ASD), attention difficulties, or sensory and language processing challenges. The system is conceived as an interactive game application designed to be simple, visually clear, and to provide constant feedback. The games feature intuitive mechanics, easily recognizable visual stimuli, and an adaptable gameplay pace, in order to facilitate the participation of young users with different levels of cognitive, motor, or communicative abilities. Through this application, the project seeks to promote learning through play, motivation, and positive reinforcement. The development of the project covers the stages of design, technical implementation, and system evaluation. During the design process, particular emphasis has been placed on usability, accessibility, and user experience criteria, following a user-centered approach. The implementation has been carried out using technologies aimed at multiplatform video game development, allowing the system to run on different devices, including mobile environments. In conclusion, the project combines digital entertainment and accessibility, which may contribute to inclusion and the development of basic skills in children with diverse cognitive, motor, and communicative profiles. Furthermore, the development of applications of this type does not require a significant additional effort from the programmer and enables this group to benefit from current technology, supporting both leisure activities and learning and educational contexts.
This Bachelor’s Final Degree Project, developed in collaboration with Pictoaplicaciones, aims to design and implement a multiplatform video game system focused on providing accessible learning and entertainment experiences, using clear and easily understandable visual language for children. Special attention is given to the inclusion of users with functional diversity and neurodivergent profiles, such as Autism Spectrum Disorder (ASD), attention difficulties, or sensory and language processing challenges. The system is conceived as an interactive game application designed to be simple, visually clear, and to provide constant feedback. The games feature intuitive mechanics, easily recognizable visual stimuli, and an adaptable gameplay pace, in order to facilitate the participation of young users with different levels of cognitive, motor, or communicative abilities. Through this application, the project seeks to promote learning through play, motivation, and positive reinforcement. The development of the project covers the stages of design, technical implementation, and system evaluation. During the design process, particular emphasis has been placed on usability, accessibility, and user experience criteria, following a user-centered approach. The implementation has been carried out using technologies aimed at multiplatform video game development, allowing the system to run on different devices, including mobile environments. In conclusion, the project combines digital entertainment and accessibility, which may contribute to inclusion and the development of basic skills in children with diverse cognitive, motor, and communicative profiles. Furthermore, the development of applications of this type does not require a significant additional effort from the programmer and enables this group to benefit from current technology, supporting both leisure activities and learning and educational contexts.
Direction
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
García García, J. Miguel (Co-tutorships)
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
García García, J. Miguel (Co-tutorships)
Court
García García, J. Miguel (Student’s tutor)
TABOADA IGLESIAS, MARÍA JESÚS (Student’s tutor)
García García, J. Miguel (Student’s tutor)
TABOADA IGLESIAS, MARÍA JESÚS (Student’s tutor)
Semidefinite optimization in polynomial optimization algorithms
Authorship
M.C.R.
Double Bachelor's Degree in Informatics Engineering and Mathematics
M.C.R.
Double Bachelor's Degree in Informatics Engineering and Mathematics
Defense date
02.12.2026 16:30
02.12.2026 16:30
Summary
This bacehlor thesis focuses on the study and application of semidefinite optimization techniques to polynomial optimization problems. It begins with the presentation of the fundamental concepts of polynomial optimization. Subsequently, the RLT technique is introduced as an algorithm for solving these types of problems, along with a practical example that illustrates its application. The following section addresses the principles of semidefinite optimization, highlighting a specific technique known as SDP cuts, which forms the basis of the computational study carried out in the final chapter. This study is conducted using the global optimization solver RAPOSa to evaluate the impact of SDP cuts on its performance. Furthermore, a sharper version of the current implementation is proposed, seeking to enhance computational efficiency.
This bacehlor thesis focuses on the study and application of semidefinite optimization techniques to polynomial optimization problems. It begins with the presentation of the fundamental concepts of polynomial optimization. Subsequently, the RLT technique is introduced as an algorithm for solving these types of problems, along with a practical example that illustrates its application. The following section addresses the principles of semidefinite optimization, highlighting a specific technique known as SDP cuts, which forms the basis of the computational study carried out in the final chapter. This study is conducted using the global optimization solver RAPOSa to evaluate the impact of SDP cuts on its performance. Furthermore, a sharper version of the current implementation is proposed, seeking to enhance computational efficiency.
Direction
GONZALEZ DIAZ, JULIO (Tutorships)
GONZALEZ RODRIGUEZ, BRAIS (Co-tutorships)
GONZALEZ DIAZ, JULIO (Tutorships)
GONZALEZ RODRIGUEZ, BRAIS (Co-tutorships)
Court
GARCIA RIO, EDUARDO (Chairman)
GARCIA LUCAS, DIEGO (Secretary)
SANCHEZ SELLERO, CESAR ANDRES (Member)
GARCIA RIO, EDUARDO (Chairman)
GARCIA LUCAS, DIEGO (Secretary)
SANCHEZ SELLERO, CESAR ANDRES (Member)
Development of an optimization functionality for an explainability module
Authorship
D.D.L.
Bachelor’s Degree in Informatics Engineering
D.D.L.
Bachelor’s Degree in Informatics Engineering
Defense date
02.20.2026 10:15
02.20.2026 10:15
Summary
This work addresses the design and implementation of an optimization functionality integrated into a corporate explainability module based on Machine Learning techniques. The starting point is an existing system that allows the interpretation of supervised models through SHAP values, with the aim of understanding the contribution of features to the prediction. However, the module lacked a mechanism to automatically identify combinations of variable values that optimize the target variable, as well as a systematic procedure to analyze complex explanatory patterns. To address these limitations, two complementary lines of work are proposed. The first focuses on the development of an interpretable clustering pipeline applied to SHAP values, integrating partition-based, hierarchical, and density-based techniques, together with automated procedures for selecting the number of clusters and dimensionality reduction strategies. This approach makes it possible to identify clusters with different performance levels, as well as groups with similar values of the target variable but different explanatory patterns, supporting interpretability through the retrieval of the original feature values and visualization outputs. The second line addresses an offline evolutionary optimization scenario in which neither an analytical fitness function nor the possibility of obtaining new real evaluations is available. To this end, a Genetic Algorithm adapted to problems with mixed variables is implemented, guided by surrogate models trained on historical data. Overall, the result is a generalizable and reusable module that combines explainability, exploratory analysis, and optimization, providing a useful tool to support data-driven decision-making in real-world environments.
This work addresses the design and implementation of an optimization functionality integrated into a corporate explainability module based on Machine Learning techniques. The starting point is an existing system that allows the interpretation of supervised models through SHAP values, with the aim of understanding the contribution of features to the prediction. However, the module lacked a mechanism to automatically identify combinations of variable values that optimize the target variable, as well as a systematic procedure to analyze complex explanatory patterns. To address these limitations, two complementary lines of work are proposed. The first focuses on the development of an interpretable clustering pipeline applied to SHAP values, integrating partition-based, hierarchical, and density-based techniques, together with automated procedures for selecting the number of clusters and dimensionality reduction strategies. This approach makes it possible to identify clusters with different performance levels, as well as groups with similar values of the target variable but different explanatory patterns, supporting interpretability through the retrieval of the original feature values and visualization outputs. The second line addresses an offline evolutionary optimization scenario in which neither an analytical fitness function nor the possibility of obtaining new real evaluations is available. To this end, a Genetic Algorithm adapted to problems with mixed variables is implemented, guided by surrogate models trained on historical data. Overall, the result is a generalizable and reusable module that combines explainability, exploratory analysis, and optimization, providing a useful tool to support data-driven decision-making in real-world environments.
Direction
MERA PEREZ, DAVID (Tutorships)
González Pintos, María (Co-tutorships)
MERA PEREZ, DAVID (Tutorships)
González Pintos, María (Co-tutorships)
Court
Barro Ameneiro, Senén (Chairman)
GARCIA FERNANDEZ, JULIAN (Secretary)
LADRA GONZALEZ, MANUEL EULOGIO (Member)
Barro Ameneiro, Senén (Chairman)
GARCIA FERNANDEZ, JULIAN (Secretary)
LADRA GONZALEZ, MANUEL EULOGIO (Member)
Design and development of a slicer for DLP 3D bioprinters
Authorship
S.G.S.
Bachelor’s Degree in Informatics Engineering
S.G.S.
Bachelor’s Degree in Informatics Engineering
Defense date
02.19.2026 10:30
02.19.2026 10:30
Summary
This Final Degree Project focuses on the design and implementation of a slicing software module intended for a DLP (Digital Light Processing) based 3D bioprinter. The developed system converts three-dimensional models (STL) into a sequence of two-dimensional layers ready for printing, integrating material management and 3D model preview functionalities. The project was implemented in Python and features an interactive graphical user interface that simplifies the configuration of printing parameters, model scaling, and process visualization. Two versions of the slicer were developed: a basic version aimed at validating the core workflow, and an optimized version that incorporates layer compression techniques. Additionally, a component separation option is included to prevent errors in multimaterial printing processes. The system was tested on the same development environment using multimaterial files provided by the co-supervisor, confirming its correct operation and stability.
This Final Degree Project focuses on the design and implementation of a slicing software module intended for a DLP (Digital Light Processing) based 3D bioprinter. The developed system converts three-dimensional models (STL) into a sequence of two-dimensional layers ready for printing, integrating material management and 3D model preview functionalities. The project was implemented in Python and features an interactive graphical user interface that simplifies the configuration of printing parameters, model scaling, and process visualization. Two versions of the slicer were developed: a basic version aimed at validating the core workflow, and an optimized version that incorporates layer compression techniques. Additionally, a component separation option is included to prevent errors in multimaterial printing processes. The system was tested on the same development environment using multimaterial files provided by the co-supervisor, confirming its correct operation and stability.
Direction
FLORES GONZALEZ, JULIAN CARLOS (Tutorships)
González Santos, Alejandro (Co-tutorships)
FLORES GONZALEZ, JULIAN CARLOS (Tutorships)
González Santos, Alejandro (Co-tutorships)
Court
VAZQUEZ CENDON, MARIA ELENA (Chairman)
GALLEGO FONTENLA, VICTOR JOSE (Secretary)
SUAREZ GAREA, JORGE ALBERTO (Member)
VAZQUEZ CENDON, MARIA ELENA (Chairman)
GALLEGO FONTENLA, VICTOR JOSE (Secretary)
SUAREZ GAREA, JORGE ALBERTO (Member)
Automatic generation of examples for querying LLMs in ontology alignment
Authorship
S.L.B.
Bachelor’s Degree in Informatics Engineering
S.L.B.
Bachelor’s Degree in Informatics Engineering
Defense date
02.19.2026 11:00
02.19.2026 11:00
Summary
Ontology Matching (OM) is a critical task for enabling data interoperability and knowledge exchange. Large Language Models (LLMs) offer new semantic capabilities to address this problem. However, the computational cost of LLMs, together with their lower efficiency and reliability, prevents their direct and exclusive use for performing ontology alignments. Some current LLM-based approaches, such as MILA, mitigate this issue by using LLMs only for the most ambiguous cases or for those that cannot be resolved with sufficient confidence. Nevertheless, the MILA system relies on a zero-shot prompting approach, which limits both the capabilities of the LLM and the amount of information provided to it. This work proposes an improvement to the MILA system by introducing dynamic examples into the prompt used to query the LLM. For each pair of concepts that MILA is unable to match with high confidence and therefore requires LLM invocation, the proposed method finds examples relevant to the evaluated pair of concepts and integrates them into the query prompt. These examples are selected dynamically from an example base composed of concept pairs that the system identifies as high-confidence alignments. In this way, the LLM receives a prompt containing successful alignment examples similar to the pair of concepts being aligned, which helps guide and contextualize the model in resolving the query. The system presented in this undergraduate thesis has improved the performance of the original MILA system, surpassing the number of LLM-required alignments that are correctly resolved in 4 out of 7 tasks and increasing the F-score by up to 10% in one of them.
Ontology Matching (OM) is a critical task for enabling data interoperability and knowledge exchange. Large Language Models (LLMs) offer new semantic capabilities to address this problem. However, the computational cost of LLMs, together with their lower efficiency and reliability, prevents their direct and exclusive use for performing ontology alignments. Some current LLM-based approaches, such as MILA, mitigate this issue by using LLMs only for the most ambiguous cases or for those that cannot be resolved with sufficient confidence. Nevertheless, the MILA system relies on a zero-shot prompting approach, which limits both the capabilities of the LLM and the amount of information provided to it. This work proposes an improvement to the MILA system by introducing dynamic examples into the prompt used to query the LLM. For each pair of concepts that MILA is unable to match with high confidence and therefore requires LLM invocation, the proposed method finds examples relevant to the evaluated pair of concepts and integrates them into the query prompt. These examples are selected dynamically from an example base composed of concept pairs that the system identifies as high-confidence alignments. In this way, the LLM receives a prompt containing successful alignment examples similar to the pair of concepts being aligned, which helps guide and contextualize the model in resolving the query. The system presented in this undergraduate thesis has improved the performance of the original MILA system, surpassing the number of LLM-required alignments that are correctly resolved in 4 out of 7 tasks and increasing the F-score by up to 10% in one of them.
Direction
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
Court
VAZQUEZ CENDON, MARIA ELENA (Chairman)
GALLEGO FONTENLA, VICTOR JOSE (Secretary)
SUAREZ GAREA, JORGE ALBERTO (Member)
VAZQUEZ CENDON, MARIA ELENA (Chairman)
GALLEGO FONTENLA, VICTOR JOSE (Secretary)
SUAREZ GAREA, JORGE ALBERTO (Member)
Acetone production plant from isopropyl alcohol
Authorship
L.L.C.
Bachelor's Degree in Chemical Engineering
L.L.C.
Bachelor's Degree in Chemical Engineering
Defense date
02.10.2026 10:00
02.10.2026 10:00
Summary
This project comprises the technical design of an acetone production plant with a 99% purity, and a capacity of 50.000 tons per year. The process is based on the catalytic dehydrogenation of isopropyl alcohol (IPA). This reaction is endothermic and takes place in a multitubular reactor, R-301, which has had a detailed mechanical design done. The reaction operates at 380ºC and 3,2 bar, and the catalyst involved is an alumina-supported metal catalyst. For the temperature control, a molten salt is used as the heat transfer fluid, selected for its excellent stability at high temperatures. The product separation has been carried out by a purification system. First, hydrogen is recovered as a high-purity sub-product by a flash separation and an absorption stage. Afterwards, the liquid stream from the separation is fed to a distillation system consisting of two columns that allow acetone to be obtained as a high-purity product and to recover the isopropanol that didn’t convert in the reactor. Moreover, this design proposes the recirculation of the retrieved azeotropic-grade IPA into the feeding stage, and the purified water to the inlet of the absorption column; which allows the optimization of the process performance. These aspects impulse the economic viability of the project, but they don’t succeed in making it profitable, as shown in the financial study included.
This project comprises the technical design of an acetone production plant with a 99% purity, and a capacity of 50.000 tons per year. The process is based on the catalytic dehydrogenation of isopropyl alcohol (IPA). This reaction is endothermic and takes place in a multitubular reactor, R-301, which has had a detailed mechanical design done. The reaction operates at 380ºC and 3,2 bar, and the catalyst involved is an alumina-supported metal catalyst. For the temperature control, a molten salt is used as the heat transfer fluid, selected for its excellent stability at high temperatures. The product separation has been carried out by a purification system. First, hydrogen is recovered as a high-purity sub-product by a flash separation and an absorption stage. Afterwards, the liquid stream from the separation is fed to a distillation system consisting of two columns that allow acetone to be obtained as a high-purity product and to recover the isopropanol that didn’t convert in the reactor. Moreover, this design proposes the recirculation of the retrieved azeotropic-grade IPA into the feeding stage, and the purified water to the inlet of the absorption column; which allows the optimization of the process performance. These aspects impulse the economic viability of the project, but they don’t succeed in making it profitable, as shown in the financial study included.
Direction
FREIRE LEIRA, MARIA SONIA (Tutorships)
González Álvarez, Julia (Co-tutorships)
FREIRE LEIRA, MARIA SONIA (Tutorships)
González Álvarez, Julia (Co-tutorships)
Court
HOSPIDO QUINTANA, ALMUDENA (Chairman)
MAURICIO IGLESIAS, MIGUEL (Secretary)
GONZALEZ GARCIA, SARA (Member)
HOSPIDO QUINTANA, ALMUDENA (Chairman)
MAURICIO IGLESIAS, MIGUEL (Secretary)
GONZALEZ GARCIA, SARA (Member)
Methanol to acetic acid production plant
Authorship
E.O.M.
Bachelor's Degree in Chemical Engineering
E.O.M.
Bachelor's Degree in Chemical Engineering
Defense date
02.10.2026 10:30
02.10.2026 10:30
Summary
Obtaining 70,000 t/year of acetic acid from methanol and carbon monoxide by means of the methanol carbonylation process. This organic compound has a wide variety of uses, from the production of cosmetic and pharmaceutical products to the food, textile and chemical industries, which makes it a product of great industrial interest. In this project, Martín Álvarez will carry out the rigorous design of the reactor, in charge of the methanol carbonylation process to obtain acetic acid, and Elena Ojea of the first column, in charge of separating acetic acid and water from the other components of the reaction medium.
Obtaining 70,000 t/year of acetic acid from methanol and carbon monoxide by means of the methanol carbonylation process. This organic compound has a wide variety of uses, from the production of cosmetic and pharmaceutical products to the food, textile and chemical industries, which makes it a product of great industrial interest. In this project, Martín Álvarez will carry out the rigorous design of the reactor, in charge of the methanol carbonylation process to obtain acetic acid, and Elena Ojea of the first column, in charge of separating acetic acid and water from the other components of the reaction medium.
Direction
FREIRE LEIRA, MARIA SONIA (Tutorships)
González Álvarez, Julia (Co-tutorships)
FREIRE LEIRA, MARIA SONIA (Tutorships)
González Álvarez, Julia (Co-tutorships)
Court
HOSPIDO QUINTANA, ALMUDENA (Chairman)
MAURICIO IGLESIAS, MIGUEL (Secretary)
GONZALEZ GARCIA, SARA (Member)
HOSPIDO QUINTANA, ALMUDENA (Chairman)
MAURICIO IGLESIAS, MIGUEL (Secretary)
GONZALEZ GARCIA, SARA (Member)
Introduction to Optimal Packing Problems
Authorship
J.M.O.C.
Double Bachelor's Degree in Informatics Engineering and Mathematics
J.M.O.C.
Double Bachelor's Degree in Informatics Engineering and Mathematics
Defense date
02.13.2026 09:45
02.13.2026 09:45
Summary
This work develops an introduction to the fundamental concepts, models, and algorithms of cutting and packing problems. To this end, a theoretical framework is initially established to unify the mathematical formulation of these problems. Following this foundation, three classic one-dimensional problems are analyzed in depth. First, the Knapsack Problem is studied, a value maximization problem for which two exact solution methods are presented, based on Dynamic Programming and Branch and Bound. Next, the Bin Packing Problem is addressed, which focuses on minimizing the use of resources. For this problem, the main heuristic algorithms are described, and an analysis of their performance is introduced using lower bounds and worst-case ratios. Thirdly, the Cutting Stock Problem is presented, a problem of great industrial relevance, whose solution is approached by obtaining the solution of the problem's linear relaxation through the column generation technique. Finally, to complete the study, a brief incursion into two-dimensional problems is made, illustrating how geometric complexity is often managed by reducing the problem to its already studied one-dimensional analogs.
This work develops an introduction to the fundamental concepts, models, and algorithms of cutting and packing problems. To this end, a theoretical framework is initially established to unify the mathematical formulation of these problems. Following this foundation, three classic one-dimensional problems are analyzed in depth. First, the Knapsack Problem is studied, a value maximization problem for which two exact solution methods are presented, based on Dynamic Programming and Branch and Bound. Next, the Bin Packing Problem is addressed, which focuses on minimizing the use of resources. For this problem, the main heuristic algorithms are described, and an analysis of their performance is introduced using lower bounds and worst-case ratios. Thirdly, the Cutting Stock Problem is presented, a problem of great industrial relevance, whose solution is approached by obtaining the solution of the problem's linear relaxation through the column generation technique. Finally, to complete the study, a brief incursion into two-dimensional problems is made, illustrating how geometric complexity is often managed by reducing the problem to its already studied one-dimensional analogs.
Direction
GONZALEZ DIAZ, JULIO (Tutorships)
GONZALEZ DIAZ, JULIO (Tutorships)
Court
VAZQUEZ CENDON, MARIA ELENA (Chairman)
LOPEZ SOMOZA, LUCIA (Secretary)
CASAS MENDEZ, BALBINA VIRGINIA (Member)
VAZQUEZ CENDON, MARIA ELENA (Chairman)
LOPEZ SOMOZA, LUCIA (Secretary)
CASAS MENDEZ, BALBINA VIRGINIA (Member)
Exploring random walks in ontology alignment with MILA
Authorship
J.M.O.C.
Double Bachelor's Degree in Informatics Engineering and Mathematics
J.M.O.C.
Double Bachelor's Degree in Informatics Engineering and Mathematics
Defense date
02.20.2026 09:15
02.20.2026 09:15
Summary
Ontology alignment identifies semantic correspondences between concepts in heterogeneous ontologies, a task that is essential for interoperability in the Semantic Web. Current approaches based on language models achieve high accuracy, but querying the model is costly for ambiguous cases that cannot be resolved through semantic similarity. This work proposes complementing semantic information with structural information using random walks. A graph is constructed that unifies the source and target ontologies, connecting them through previously identified high-confidence correspondences. The hypothesis is that two concepts from different ontologies are likely equivalent if a random walker starting from one frequently reaches the other, indicating proximity within the graph structure. Two algorithms (Biased Random Walks and Random Walks with Restart) were implemented, integrated into the MILA system, and evaluated on five tasks from OAEI 2024: three from the Bio-ML track, one from the biodiversity domain, and one from the anatomy domain. The results show that structural information adds value selectively: it improves the candidate ranking for certain ontology pairs where semantic similarity is weakly discriminative, but it can introduce noise when semantic signals are already strong. Advisor mode (re-ranking) is more robust than predictor mode (autonomous decision-making), and Biased Random Walks outperform Random Walks with Restart. We conclude that random walks are a complement -not a substitute- for semantic techniques, and that their benefit depends on the structural characteristics of the ontology pair.
Ontology alignment identifies semantic correspondences between concepts in heterogeneous ontologies, a task that is essential for interoperability in the Semantic Web. Current approaches based on language models achieve high accuracy, but querying the model is costly for ambiguous cases that cannot be resolved through semantic similarity. This work proposes complementing semantic information with structural information using random walks. A graph is constructed that unifies the source and target ontologies, connecting them through previously identified high-confidence correspondences. The hypothesis is that two concepts from different ontologies are likely equivalent if a random walker starting from one frequently reaches the other, indicating proximity within the graph structure. Two algorithms (Biased Random Walks and Random Walks with Restart) were implemented, integrated into the MILA system, and evaluated on five tasks from OAEI 2024: three from the Bio-ML track, one from the biodiversity domain, and one from the anatomy domain. The results show that structural information adds value selectively: it improves the candidate ranking for certain ontology pairs where semantic similarity is weakly discriminative, but it can introduce noise when semantic signals are already strong. Advisor mode (re-ranking) is more robust than predictor mode (autonomous decision-making), and Biased Random Walks outperform Random Walks with Restart. We conclude that random walks are a complement -not a substitute- for semantic techniques, and that their benefit depends on the structural characteristics of the ontology pair.
Direction
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
Court
Barro Ameneiro, Senén (Chairman)
GARCIA FERNANDEZ, JULIAN (Secretary)
LADRA GONZALEZ, MANUEL EULOGIO (Member)
Barro Ameneiro, Senén (Chairman)
GARCIA FERNANDEZ, JULIAN (Secretary)
LADRA GONZALEZ, MANUEL EULOGIO (Member)
Vinyl acetate (monomer) production plant
Authorship
A.P.D.
Bachelor's Degree in Chemical Engineering
A.P.D.
Bachelor's Degree in Chemical Engineering
Defense date
02.10.2026 10:00
02.10.2026 10:00
Summary
This project consists of the development of a vinyl acetate (monomer) production plant. This chemical product is in great demand at industrial level due to all the uses it has. It is mainly used for the creation of polymers such as PVA or PVP. For its production, ethylene, acetic acid and oxygen were used as raw materials. The process is characterized by the fact that it is carried out in the gas phase. The project was carried out by students Darío Senín Sotelo and Alberto Pisco Domínguez, who designed, respectively, the R-201 multitubular heterogeneous catalytic reactor and the T-302 physical absorption column.
This project consists of the development of a vinyl acetate (monomer) production plant. This chemical product is in great demand at industrial level due to all the uses it has. It is mainly used for the creation of polymers such as PVA or PVP. For its production, ethylene, acetic acid and oxygen were used as raw materials. The process is characterized by the fact that it is carried out in the gas phase. The project was carried out by students Darío Senín Sotelo and Alberto Pisco Domínguez, who designed, respectively, the R-201 multitubular heterogeneous catalytic reactor and the T-302 physical absorption column.
Direction
Rodríguez Figueiras, Óscar (Tutorships)
Rodríguez Figueiras, Óscar (Tutorships)
Court
CARBALLA ARCOS, MARTA (Chairman)
FRANCO RUIZ, DANIEL JOSE (Secretary)
EIBES GONZALEZ, GEMMA MARIA (Member)
CARBALLA ARCOS, MARTA (Chairman)
FRANCO RUIZ, DANIEL JOSE (Secretary)
EIBES GONZALEZ, GEMMA MARIA (Member)
Domain adaptation based on deep learning networks for multispectral image classification
Authorship
M.G.P.P.
Bachelor’s Degree in Informatics Engineering
M.G.P.P.
Bachelor’s Degree in Informatics Engineering
Defense date
02.19.2026 17:00
02.19.2026 17:00
Summary
In the field of remote-sensing, the use of deep learning plays a very important role in classification and segmentation tasks, enabling the efficient detection and grouping of materials according to the class to which they belong, this being a major advantage over manual labeling. However, training classification models faces a very common problem in reality: domain shift. This phenomenon occurs when a model is trained on data whose distribution differs from that of the data which has been infered. As a solution to this problem, this work proposes an implementation based on the model Cycle Consistent Adversarial Domain Adaptation (CyCADA), which, unlike traditional adaptation methods, combines pixel-level image translation through adversarial adaptation with semantic consistency. Its goal is to adapt a model from a source domain to his target domain without using labels in the latter, ensuring that the transformations preserve both the visual appearance and the structural consistency of the objects in the image. To optimize the adaptation, a set of experiments was carried out to determine the best CyCADA architecture compatible with our dataset. Besides optimizing this architecture, with the goal of comparing its effect against not using adaptation, other adaptation techniques were tested to allow comparison of results across different methods like traditional techniques (CORAL, invariant band selection, and pseudolabeling). The results of these experiments led to the conclusion that, although the proposed implementation does improve classification performance after adaptation, compared to other techniques, the improvements are not significant enough considering the computational cost it presents. Therefore it was concluded that, for this scenario, CyCADA is not the best solution for adapting the source domain to the target. Among the solutions that showed benefits over classification without adaptation, stands out the selection of invariant feature and self-labeling with segmentation.
In the field of remote-sensing, the use of deep learning plays a very important role in classification and segmentation tasks, enabling the efficient detection and grouping of materials according to the class to which they belong, this being a major advantage over manual labeling. However, training classification models faces a very common problem in reality: domain shift. This phenomenon occurs when a model is trained on data whose distribution differs from that of the data which has been infered. As a solution to this problem, this work proposes an implementation based on the model Cycle Consistent Adversarial Domain Adaptation (CyCADA), which, unlike traditional adaptation methods, combines pixel-level image translation through adversarial adaptation with semantic consistency. Its goal is to adapt a model from a source domain to his target domain without using labels in the latter, ensuring that the transformations preserve both the visual appearance and the structural consistency of the objects in the image. To optimize the adaptation, a set of experiments was carried out to determine the best CyCADA architecture compatible with our dataset. Besides optimizing this architecture, with the goal of comparing its effect against not using adaptation, other adaptation techniques were tested to allow comparison of results across different methods like traditional techniques (CORAL, invariant band selection, and pseudolabeling). The results of these experiments led to the conclusion that, although the proposed implementation does improve classification performance after adaptation, compared to other techniques, the improvements are not significant enough considering the computational cost it presents. Therefore it was concluded that, for this scenario, CyCADA is not the best solution for adapting the source domain to the target. Among the solutions that showed benefits over classification without adaptation, stands out the selection of invariant feature and self-labeling with segmentation.
Direction
SUAREZ GAREA, JORGE ALBERTO (Tutorships)
QUESADA BARRIUSO, PABLO (Co-tutorships)
SUAREZ GAREA, JORGE ALBERTO (Tutorships)
QUESADA BARRIUSO, PABLO (Co-tutorships)
Court
QUESADA BARRIUSO, PABLO (Student’s tutor)
SUAREZ GAREA, JORGE ALBERTO (Student’s tutor)
QUESADA BARRIUSO, PABLO (Student’s tutor)
SUAREZ GAREA, JORGE ALBERTO (Student’s tutor)
Efficient querying of geospatial networks: A comparison between Object-Relational and Graph Databases.
Authorship
D.R.R.
Bachelor’s Degree in Informatics Engineering
D.R.R.
Bachelor’s Degree in Informatics Engineering
Defense date
02.19.2026 10:00
02.19.2026 10:00
Summary
Efficient management of public transportation networks presents increasingly complex challenges due to the need to integrate large volumes of both spatial and temporal data. Consequently, it becomes necessary to evaluate whether traditional SQL architectures are still optimal, or, alternatively, graph-oriented paradigms offer substantial advantages. In this work, the geospatial networks have been modeled in two different DBMS, PostgreSQL and Neo4J, following the GTFS standard. Subsequently, a set of representative queries, which answer real information needs from both users and service companies, have been developed. Furthermore, the retrieved data have been visualized using QGIS and Folium, and finally, the features of both systems have been assessed, emphasizing not only computational efficiency, but also ease of modelling and query expression. The results show that PostgreSQL offers superior performance in terms of data ingestion and large-scale aggregation, while also providing excellent support for geospatial data. On the other hand, Neo4J excels at queries that require traversal from a small number of initial nodes, and offers a more readable, easy-to-maintain syntax, although with limitations in library support.
Efficient management of public transportation networks presents increasingly complex challenges due to the need to integrate large volumes of both spatial and temporal data. Consequently, it becomes necessary to evaluate whether traditional SQL architectures are still optimal, or, alternatively, graph-oriented paradigms offer substantial advantages. In this work, the geospatial networks have been modeled in two different DBMS, PostgreSQL and Neo4J, following the GTFS standard. Subsequently, a set of representative queries, which answer real information needs from both users and service companies, have been developed. Furthermore, the retrieved data have been visualized using QGIS and Folium, and finally, the features of both systems have been assessed, emphasizing not only computational efficiency, but also ease of modelling and query expression. The results show that PostgreSQL offers superior performance in terms of data ingestion and large-scale aggregation, while also providing excellent support for geospatial data. On the other hand, Neo4J excels at queries that require traversal from a small number of initial nodes, and offers a more readable, easy-to-maintain syntax, although with limitations in library support.
Direction
RIOS VIQUEIRA, JOSE RAMON (Tutorships)
RIOS VIQUEIRA, JOSE RAMON (Tutorships)
Court
VAZQUEZ CENDON, MARIA ELENA (Chairman)
GALLEGO FONTENLA, VICTOR JOSE (Secretary)
SUAREZ GAREA, JORGE ALBERTO (Member)
VAZQUEZ CENDON, MARIA ELENA (Chairman)
GALLEGO FONTENLA, VICTOR JOSE (Secretary)
SUAREZ GAREA, JORGE ALBERTO (Member)
Acrylic acid production plant using propylene oxidation
Authorship
S.R.S.
Bachelor's Degree in Chemical Engineering
S.R.S.
Bachelor's Degree in Chemical Engineering
Defense date
02.10.2026 09:15
02.10.2026 09:15
Summary
This Final Degree Project consists of the design of an industrial plant for the production of 20,000 tonnes per year of acrylic acid through the catalytic oxidation of propylene, operating continuously. The process includes a feedstock conditioning stage, where the feed streams (propylene and air) are prepared and utilities such as nitrogen are managed to ensure safe and stable operating conditions. Next, the reaction stage is carried out in two reactors in series, with acrolein formed as an intermediate product, this configuration being selected due to its greater control, selectivity and performance. After the reaction, the stream undergoes cooling and an initial separation through a flash and absorption, removing gases and obtaining an aqueous liquid stream of crude product. Finally, in the purification stage, impurities such as water and acetic acid are removed through separation operations (including distillation), achieving the acrylic acid quality specifications and enabling recovery of the by-product. The project includes material and energy balances, description and basic sizing of equipment, control and instrumentation, industrial safety, and technical and economic feasibility assessment. The rigorous design of the reactor responsible for the first reaction stage of selective oxidation of propylene to acrolein was carried out by Mateo Segade Abelleira. The rigorous design of the distillation column responsible for purifying the product and the by-product of the plant was carried out by Sergio Rial Suárez.
This Final Degree Project consists of the design of an industrial plant for the production of 20,000 tonnes per year of acrylic acid through the catalytic oxidation of propylene, operating continuously. The process includes a feedstock conditioning stage, where the feed streams (propylene and air) are prepared and utilities such as nitrogen are managed to ensure safe and stable operating conditions. Next, the reaction stage is carried out in two reactors in series, with acrolein formed as an intermediate product, this configuration being selected due to its greater control, selectivity and performance. After the reaction, the stream undergoes cooling and an initial separation through a flash and absorption, removing gases and obtaining an aqueous liquid stream of crude product. Finally, in the purification stage, impurities such as water and acetic acid are removed through separation operations (including distillation), achieving the acrylic acid quality specifications and enabling recovery of the by-product. The project includes material and energy balances, description and basic sizing of equipment, control and instrumentation, industrial safety, and technical and economic feasibility assessment. The rigorous design of the reactor responsible for the first reaction stage of selective oxidation of propylene to acrolein was carried out by Mateo Segade Abelleira. The rigorous design of the distillation column responsible for purifying the product and the by-product of the plant was carried out by Sergio Rial Suárez.
Direction
RODIL RODRIGUEZ, EVA (Tutorships)
RODIL RODRIGUEZ, EVA (Tutorships)
Court
González Álvarez, Julia (Chairman)
GOMEZ DIAZ, DIEGO (Secretary)
GIL GONZALEZ, ALVARO (Member)
González Álvarez, Julia (Chairman)
GOMEZ DIAZ, DIEGO (Secretary)
GIL GONZALEZ, ALVARO (Member)
Acrylic acid production plant using propylene oxidation
Authorship
M.S.A.
Bachelor's Degree in Chemical Engineering
M.S.A.
Bachelor's Degree in Chemical Engineering
Defense date
02.10.2026 09:15
02.10.2026 09:15
Summary
This Final Degree Project consists of the design of an industrial plant for the production of 20,000 tonnes per year of acrylic acid through the catalytic oxidation of propylene, operating continuously. The process includes a feedstock conditioning stage, where the feed streams (propylene and air) are prepared and utilities such as nitrogen are managed to ensure safe and stable operating conditions. Next, the reaction stage is carried out in two reactors in series, with acrolein formed as an intermediate product, this configuration being selected due to its greater control, selectivity and performance. After the reaction, the stream undergoes cooling and an initial separation through a flash and absorption, removing gases and obtaining an aqueous liquid stream of crude product. Finally, in the purification stage, impurities such as water and acetic acid are removed through separation operations (including distillation), achieving the acrylic acid quality specifications and enabling recovery of the by-product. The project includes material and energy balances, description and basic sizing of equipment, control and instrumentation, industrial safety, and technical and economic feasibility assessment. The rigorous design of the reactor responsible for the first reaction stage of selective oxidation of propylene to acrolein was carried out by Mateo Segade Abelleira. The rigorous design of the distillation column responsible for purifying the product and the by-product of the plant was carried out by Sergio Rial Suarez.
This Final Degree Project consists of the design of an industrial plant for the production of 20,000 tonnes per year of acrylic acid through the catalytic oxidation of propylene, operating continuously. The process includes a feedstock conditioning stage, where the feed streams (propylene and air) are prepared and utilities such as nitrogen are managed to ensure safe and stable operating conditions. Next, the reaction stage is carried out in two reactors in series, with acrolein formed as an intermediate product, this configuration being selected due to its greater control, selectivity and performance. After the reaction, the stream undergoes cooling and an initial separation through a flash and absorption, removing gases and obtaining an aqueous liquid stream of crude product. Finally, in the purification stage, impurities such as water and acetic acid are removed through separation operations (including distillation), achieving the acrylic acid quality specifications and enabling recovery of the by-product. The project includes material and energy balances, description and basic sizing of equipment, control and instrumentation, industrial safety, and technical and economic feasibility assessment. The rigorous design of the reactor responsible for the first reaction stage of selective oxidation of propylene to acrolein was carried out by Mateo Segade Abelleira. The rigorous design of the distillation column responsible for purifying the product and the by-product of the plant was carried out by Sergio Rial Suarez.
Direction
RODIL RODRIGUEZ, EVA (Tutorships)
RODIL RODRIGUEZ, EVA (Tutorships)
Court
González Álvarez, Julia (Chairman)
GOMEZ DIAZ, DIEGO (Secretary)
GIL GONZALEZ, ALVARO (Member)
González Álvarez, Julia (Chairman)
GOMEZ DIAZ, DIEGO (Secretary)
GIL GONZALEZ, ALVARO (Member)
PICTOESCENAS: an application to improve communication and comprehension in people with oral language disorders
Authorship
D.T.D.
Bachelor’s Degree in Informatics Engineering
D.T.D.
Bachelor’s Degree in Informatics Engineering
Defense date
02.18.2026 09:00
02.18.2026 09:00
Summary
A web application for creating interactive educational scenes based on images and pictograms, aimed at people with oral language disorders such as Autism Spectrum Disorder (ASD). The tool allows defining interactive areas on an image and associating events such as text, sounds or sentences automatically translated into pictograms to enhance communication and comprehension.
A web application for creating interactive educational scenes based on images and pictograms, aimed at people with oral language disorders such as Autism Spectrum Disorder (ASD). The tool allows defining interactive areas on an image and associating events such as text, sounds or sentences automatically translated into pictograms to enhance communication and comprehension.
Direction
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
García García, J. Miguel (Co-tutorships)
TABOADA IGLESIAS, MARÍA JESÚS (Tutorships)
García García, J. Miguel (Co-tutorships)
Court
García García, J. Miguel (Student’s tutor)
TABOADA IGLESIAS, MARÍA JESÚS (Student’s tutor)
García García, J. Miguel (Student’s tutor)
TABOADA IGLESIAS, MARÍA JESÚS (Student’s tutor)
Butadiene production plant from n-butane catalytic dehydrogenation
Authorship
A.Y.R.
Bachelor's Degree in Chemical Engineering
A.Y.R.
Bachelor's Degree in Chemical Engineering
Defense date
02.10.2026 12:30
02.10.2026 12:30
Summary
The objective of this project is the design, at the basic engineering level, of a 50,000 tons per year production plant of 1,3-butadiene at 99.99% purity from the catalytic dehydrogenation of n-butane with a continuous operation mode of 24 hours per day for 330 days per year, reserving the remaining days for plant maintenance. It includes a detailed design of the 1-butene oxidative dehydrogenation reactor (R202) carried out by the student Alejandra Novo Iñigo, of the plate distillation column (T-304) by Andrea Yáñez Rey and of the absorption tower (T-301) by Celia Martínez Fernández. This project is presented as the Final Degree Project of the students Alejandra Novo Íñigo, Celia Martínez Fernández and Andrea Yáñez Rey with the aim of obtaining the degree in Chemical Engineering awarded by the University of Santiago de Compostela.
The objective of this project is the design, at the basic engineering level, of a 50,000 tons per year production plant of 1,3-butadiene at 99.99% purity from the catalytic dehydrogenation of n-butane with a continuous operation mode of 24 hours per day for 330 days per year, reserving the remaining days for plant maintenance. It includes a detailed design of the 1-butene oxidative dehydrogenation reactor (R202) carried out by the student Alejandra Novo Iñigo, of the plate distillation column (T-304) by Andrea Yáñez Rey and of the absorption tower (T-301) by Celia Martínez Fernández. This project is presented as the Final Degree Project of the students Alejandra Novo Íñigo, Celia Martínez Fernández and Andrea Yáñez Rey with the aim of obtaining the degree in Chemical Engineering awarded by the University of Santiago de Compostela.
Direction
RODIL RODRIGUEZ, EVA (Tutorships)
RODIL RODRIGUEZ, EVA (Tutorships)
Court
Omil Prieto, Francisco (Chairman)
Rodríguez Figueiras, Óscar (Secretary)
VAL DEL RIO, MARIA ANGELES (Member)
Omil Prieto, Francisco (Chairman)
Rodríguez Figueiras, Óscar (Secretary)
VAL DEL RIO, MARIA ANGELES (Member)