Public display of deposited theses

DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE

• FERRANDO MONSONÍS, JAVIER: Interpretability in Natural Language Processing and Machine Translation
  
  Author: FERRANDO MONSONÍS, JAVIER
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE
  Department: Department of Computer Science (CS)
  Mode: Normal
  Deposit date: 13/11/2025
  Deposit END date: 26/11/2025
  Thesis director: RUIZ COSTA-JUSSA, MARTA
  Thesis abstract: This thesis presents a set of methods and analyses designed to improve our understanding of the internal mechanisms of Transformer-based models in natural language processing and machine translation.This work first investigates the role of attention weights in encoder-decoder Transformers, showing that while they do not provide accurate word alignments, they nonetheless help explain model predictions and contribute to a deeper understanding of translation quality.A central contribution of the dissertation is the development of ALTI and its extensions, which offer a new approach to input attribution. These methods challenge prior assumptions about the explanatory power of attention mechanisms and reveal how information propagates between encoder and decoder components. In doing so, they also shed light on sources of hallucinations in translation systems.Further, the thesis introduces techniques to attribute predictions to individual components and positions, enabling contrastive explanations of linguistic behavior. These explanations clarify how language models represent and solve different linguistic phenomena.The dissertation also proposes a methodology for tracking information flow during inference, offering insight into how various components contribute to model predictions. This allows for the identification of domain-specialized components and a better understanding of how representations are transformed across layers.Lastly, the analysis of cross-lingual circuit similarities reveals shared structural patterns in how models handle different languages. These findings point to potential universal mechanisms in language models.Collectively, this thesis advances the interpretability of Transformer models by providing tools and frameworks for probing, attributing, and understanding the behavior of complex NLP systems.
  

DOCTORAL DEGREE IN AUTOMATIC CONTROL, ROBOTICS AND VISION

• SHEIKHSAMAD, MOHAMMAD: Learning Methods in Planning and Control for Autonomous Vehicles and Robotic Manipulation
  
  Author: SHEIKHSAMAD, MOHAMMAD
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN AUTOMATIC CONTROL, ROBOTICS AND VISION
  Department: Institute of Industrial and Control Engineering (IOC)
  Mode: Normal
  Deposit date: 11/11/2025
  Deposit END date: 24/11/2025
  Thesis director: SUAREZ FEIJOO, RAUL | ROSELL GRATACOS, JOAN
  Thesis abstract: This thesis deals with the application of machine learning (ML) and deep learning (DL) techniques to enhance planning and control tasks across different fields, particularly on autonomous vehicles and robotic hands. Specifically, it addresses the learning-based development of a robust trajectory-tracking controller for autonomous vehicles, an adaptive path planner for robotic hands enabling dexterous manipulation, and a human-in-the-loop controller for myoelectric robotic hands to ensure precise grasp force regulation.In the field of autonomous vehicles, the thesis develops a Takagi–Sugeno (TS) controller using the Adaptive Neuro-Fuzzy Inference System (ANFIS) as a learning-based approach to infer a control strategy from input–output data of an existing controller. The closed-loop stability of the system is analyzed using Lyapunov theory and Linear Matrix Inequalities (LMIs). The proposed controller eliminates the need for online optimization, significantly reduces computational cost, and enhances real-time performance. Its effectiveness is validated through simulations on a small-scale autonomous vehicle.In the field of robotic dexterous manipulation, the thesis introduces three learning-based path planners using ANFIS and Deep Neural Networks (DNNs) to learn heuristics from an analytical planner and self-tune their parameters based on prior experience. This approach enables robots to manipulate objects of varying shapes, sizes, and material properties. The proposed planners are validated through real-world experiments using an Allegro robotic hand, demonstrating robustness against sensor noise and environmental disturbances.In the field of robotic grasping, the thesis presents a myocontrolled human-in-the-loop (HITL) system for precise grasp strength regulation. The system integrates both DNN-based and fuzzy-based force controllers. The fuzzy-based controller leverages fuzzy logic, with parameter optimization guided by user preferences collected through a graphical user interface (GUI) using Global Learning of Input–Output Strategies from Pairwise Preferences (GLISp). These controllers are compared against heuristic model-based controllers, and the system is validated through real-world experiments using the AR10 robotic hand, showing enhanced adaptability and fine-grained force regulation capabilities.The findings of this research contribute to the advancement of intelligent planning and control systems across multiple application areas, paving the way for more efficient, adaptive, and stable automation in real-world scenarios.
  

DOCTORAL DEGREE IN COMPUTATIONAL AND APPLIED PHYSICS

• MIRZAY SHAHIM, MAAHIN: Catalytic Properties of Amorphous Alloys
  
  Author: MIRZAY SHAHIM, MAAHIN
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN COMPUTATIONAL AND APPLIED PHYSICS
  Department: Department of Physics (FIS)
  Mode: Normal
  Deposit date: 17/11/2025
  Deposit END date: 28/11/2025
  Thesis director: PINEDA SOLER, ELOY | SOLER TURU, LLUIS
  Thesis abstract: This thesis explores the catalytic potential of metallic glasses (MGs) and their combination with cerium oxide (CeO₂) for low-temperature carbon monoxide (CO) oxidation and CO preferential oxidation (COPrOx) reactions. Metallic glasses, due to their non-crystalline structure and tunable composition, offer a promising platform for catalytic applications when appropriately engineered. The study focuses on three primary MG systems: Ce65Al35, Pd77Si16.5Cu6.5, and Cu48Zr48Al4, examining their structural characteristics, and catalytic behavior. The results showed that the Ce65Al35 metallic glass has limited catalytic activity, even after various activation treatments such as ball milling, calcination, or combination with CeO2. However, doping the binary Ce-Al system with Pd (Ce61Al35Pd4) markedly improved performance, achieving 100% CO conversion at 300°C when ball milled. Interestingly, mixing this ternary MG with CeO₂ did not provide further enhancement, indicating that Pd’s role is dominant and not synergistic with ceria. The Pd77Si16.5Cu6.5 MG emerged as the most effective standalone catalyst, delivering full CO conversion at only 240°C. Which could be attributed to Pd and its optimized distribution in the amorphous matrix. Control experiments with binary alloys (Pd77Si23 and Cu6Si94) highlighted the importance of both composition and structural processing, particularly the necessity of melt spinning and ball milling to generate active, fine-particle structures.Another major contribution of this work is the development and detailed characterization of Cu-based MG/CeO2 composites, especially Cu48Zr48Al4.These systems showed strong activity and stability in both CO and COPrOx reactions, with performance enhanced through ball milling. Structural and operando analyses (XPS, EXAFS, NEXAFS, and XRD) confirmed that the catalysts undergo surface rearrangement during reaction, stabilizing catalytically active Cu(I) atoms. A spontaneous aging phenomenon and a similar change under hydrogen pre-reduction pointed to the dynamic evolution of active sites during real operation conditions. This study demonstrates that mechanochemical synthesis and careful structural design of MG/CeO₂ composites enable the development of efficient, low-cost, and stable oxidation catalysts. These findings offer new strategies for creating highly active materials for pollution control and hydrogen purification technologies, opening the path to use amorphous metals for heterogeneous catalysis.
  

DOCTORAL DEGREE IN COMPUTER ARCHITECTURE

• BANCHELLI GRACIA, FABIO FRANCISCO: Evaluation and methods to increase efficiency of HPC systems with different maturity levels
  
  Author: BANCHELLI GRACIA, FABIO FRANCISCO
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN COMPUTER ARCHITECTURE
  Department: Department of Computer Architecture (DAC)
  Mode: Normal
  Deposit date: 12/11/2025
  Deposit END date: 25/11/2025
  Thesis director: MANTOVANI, FILIPPO | GARCIA GASULLA, MARTA
  Thesis abstract: High-Performance Computing (HPC) has entered an era of increasing architectural diversity and complexity, with systems ranging from experimental prototypes to large-scale production machines. This evolution presents a fundamental challenge: how to consistently evaluate performance, scalability, and efficiency across platforms with varying levels of technological maturity. Traditional benchmarking methods, while effective for fully deployed systems, often fall short when applied to early-stage prototypes where software stacks are incomplete or hardware is still under development.This thesis proposes and develops a comprehensive evaluation methodology capable of addressing these challenges. The approach gives a multi-layered perspective on performance, and it is structured around three complementary levels: micro-benchmarks, standard HPC benchmarks, and full scientific applications. Technology Readiness Levels (TRLs) are introduced as a guiding concept, allowing the methodology to be adapted according to the maturity of the system under study. At high TRL, the methodology enables comparative assessments of production supercomputers, while at low TRL, it helps identify bottlenecks and optimization opportunities early in the design cycle.The thesis contributes both conceptual and practical tools. It formalizes performance and efficiency models (including Roofline, Top-Down, and efficiency metrics) and demonstrates their use across multiple architectures. It further extends tracing and monitoring capabilities for emerging processors, introduces methods to access and interpret hardware counters on novel architectures such as \riscv, and evaluates the integration of experimental hardware through Software Development Vehicles (SDVs) and FPGA-based emulation. These tools are validated through case studies on production systems, such as the MareNostrum 5 supercomputer and other HPC clusters deployed at the Barcelona Supercomputing Center (BSC), as well as on prototypes from European projects, such as EPAC.Results show that the proposed methodology provides actionable insights at all maturity levels: from guiding hardware-software co-design in early-stage processors to enabling reproducible performance comparisons across pre-exascale systems. Beyond benchmarking, it provides valuable feedback for hardware architects, system software developers, and application scientists alike. By bridging the gap between low-TRL prototypes and production-ready HPC systems, this work contributes to building a consistent framework for evaluating and improving the efficiency of future European and global supercomputers.
  
• BARRERA HERRERA, JAVIER ENRIQUE: Improving Time Predictability and Code Coverage of Embedded GPUs for Real-Time Systems
  
  Author: BARRERA HERRERA, JAVIER ENRIQUE
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN COMPUTER ARCHITECTURE
  Department: Department of Computer Architecture (DAC)
  Mode: Normal
  Deposit date: 07/11/2025
  Deposit END date: 20/11/2025
  Thesis director: CAZORLA ALMEIDA, FRANCISCO JAVIER | KOSMIDIS, LEONIDAS
  Thesis abstract: This dissertation addresses challenges that the adoption of GPUs in Critical Embedded Systems (CES) faces, namely, Time Predictability and Code Coverage. Different domains that deploy CES are constantly adding Artificial Intelligence (AI)-based features, such as autonomous driving, that demand high performance levels. Multi-Processors Sytem-on-Chip (MPSoCs) are widely used to provide said performance levels, as they are equipped with accelerators, among which, Graphics Processing Units (GPUs) are a common choice. However, CES must undergo a rigorous Verification and Validation (V&V) process, in which a certain level of Execution Time Determinism (ETD) must be guaranteed. The use of several tasks to increase the overall utilization introduces contention in shared resources, which induces time variability. To provide the ETD guarantees, the time variability must be either mitigated or tracked and controlled. Another challenge for the adoption of GPUs in CES, is that the V&V process demands evidence of the thoroughness of the testing phase, for which Code Coverage is used as a test quality indicator. However, Code Coverage, as traditionally understood for sequential CES does not cover all possible scenarios in which a GPU thread might execute.For low-criticality and mixed-criticality CES, we contend that we can allow tasks to share the Last Level Cache (LLC) if hardware support for contention tracking is provided. Providing a clear understanding on how tasks contend with each other enables CES developers to balance performance and time predictability. For high-criticality CES, it is a common practice to implement LLC partitioning as it allows tasks to access LLC without suffering from inter-kernel contention, however, tasks may experience a performance loss due to lack of resources. In this Thesis, we propose Demotion Counters, a novel technique that tightly tracks how much each task has been demoted towards eviction in the LLC, thus, effectively quantifying their impact in CES. Additionally, we also assess the use of NVIDIA’s Multi-Instance GPU (MIG) feature as means to improve ETD in high-criticality CES.Code Coverage is used as a test quality indicator to provide evidence of the thoroughness of the testing, as required by the V&V process. However, if applied as traditionally understood, it will ignore the threading dimension of GPUs. Threads have private regions of memory, as well as shared regions at different granularities. This means that errors that are innocuous to one thread are potentially harmful for another, hence, it does not cover all possible cases under which GPU threads might execute. In this Thesis, we propose the use of Per-Thread Statement Coverage (PTSC), which tracks the Code Coverage at thread granularity. In order to mitigate the overheads caused by PTSC, several variants that apply different orthogonal optimizations are also proposed. Finally, we also evaluate the potential benefits of using hardware support for PTSC, mitigating the memory consumption of PTSC, as well as the execution time impact at deployment.Summarizing, this Thesis advances the state of the art in the adoption of GPUs in CES. The proposal of hardware contention tracking support and assessment of NVIDIA’s MIG, as means to improve ETD, effectively tackles the Time Predictability challenge in shared LLC. The proposal of software PTSC allows providing CES designers with the whole picture of the execution in commercially available GPUs. The use of hardware support for PTSC mitigates the overheads of software PTSC in deployment, while the different compression techniques reduce the volume of information during testing phase without losing data. Therefore, this Thesis provides means to face the Time Predictability and Code Coverage challenges of GPUs in CES.
  
• SABRI ABREBEKOH, MOHAMMAD: Improving Efficiency of ReRAM-Based Accelerators for Cognitive Computing Workloads
  
  Author: SABRI ABREBEKOH, MOHAMMAD
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN COMPUTER ARCHITECTURE
  Department: Department of Computer Architecture (DAC)
  Mode: Normal
  Deposit date: 07/11/2025
  Deposit END date: 20/11/2025
  Thesis director: GONZÁLEZ COLÁS, ANTONIO MARIA | RIERA VILLANUEVA, MARC
  Thesis abstract: Deep Neural Networks (DNNs) have achieved remarkable success across a wide range of applications. The main operation in DNNs is the dot product between quantized input activations and weights. Previous works have proposed memory-centric architectures based on the Processing-in-Memory (PuM) paradigm. ReRAM technology is especially appealing for PuM-based DNN accelerators because of its high density for weight storage, low leakage energy, low read latency, and high-performance capabilities to perform DNN dot products massively in parallel within ReRAM crossbars. However, there are three main bottlenecks in ReRAM-based accelerators.First, the energy-hungry Analog-to-Digital Converter (ADC) required for in-ReRAM analog computations, which undermines the efficiency and performance benefits of PuM. To improve energy efficiency, we present ReDy, a hardware accelerator that implements a novel ReRAM-centric dynamic quantization scheme, leveraging bit-serial streaming and processing of activations. The energy consumption of ReRAM-based DNN accelerators is directly proportional to the numerical precision of input activations in each layer. ReDy exploits the fact that activations in convolutional layers are often grouped according to filter sizes and crossbar dimensions. It quantizes each group of activations on-the-fly with different precision levels, based on a heuristic that considers the statistical distribution of each group. Overall, ReDy significantly reduces ReRAM crossbar activity and the number of A/D conversions compared to static 8-bit uniform quantization. Evaluated on a set of modern CNNs, ReDy achieves on average 13% energy savings over an ISAAC-like accelerator, with negligible area overhead.Second, the costly writing process of ReRAM cells has led to accelerators designed with enough crossbar capacity to store entire DNN models. Given the continuous growth of DNN model sizes, this approach is infeasible for some networks and inefficient due to huge hardware requirements. These accelerators lack flexibility and face an adaptability challenge. To address this, we introduce ARAS, a cost-effective ReRAM-based accelerator that uses a smart scheduler to adapt various DNNs to resource-limited hardware. ARAS also overlaps computation of one layer with weight writing of others to mitigate high ReRAM write latency. Furthermore, ARAS introduces optimizations to reduce the energy overhead of ReRAM writes, including re-encoding weights to increase similarity across layers and reduce energy when overwriting cells. Overall, ARAS significantly reduces ReRAM write activity. Evaluated on multiple DNN models, ARAS delivers up to 2.2× speedup and 45% energy savings compared to a baseline PuM accelerator without optimizations, and up to 1.5× speedup and 62% energy savings compared to a TPU-like accelerator.Third, ReRAM cells suffer from limited endurance due to wear-out caused by repeated updates during inference, reducing the lifespan of ReRAM-based accelerators. Overcoming this endurance limitation is essential for making such accelerators viable in long-term, high-performance DNN inference. To address this, we propose Hamun, an approximate computing method designed to extend the lifespan of ReRAM-based accelerators through multiple optimizations. Hamun introduces a mechanism to detect and retire faulty cells caused by wear-out, preventing them from degrading accuracy. It also applies wear-leveling techniques across different abstraction levels and introduces a batch execution scheme to maximize cell utilization across inferences. Additionally, Hamun leverages the fault-tolerance of DNNs with a new approximation method that delays cell retirement, reducing the performance penalty and further extending lifespan. Evaluated on a set of DNNs, Hamun improves lifespan by 13.2× over a state-of-the-art baseline, with its main contributions coming from fault handling (4.6×) and batch execution (2.6×).
  

DOCTORAL DEGREE IN CONSTRUCTION ENGINEERING

• POSADA CÁRCAMO, HÉCTOR JOSÉ: Digital Twins for Concrete Building Construction Processes
  
  Author: POSADA CÁRCAMO, HÉCTOR JOSÉ
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN CONSTRUCTION ENGINEERING
  Department: Department of Civil and Environmental Engineering (DECA)
  Mode: Normal
  Deposit date: 12/11/2025
  Deposit END date: 25/11/2025
  Thesis director: CHACÓN FLORES, ROLANDO ANTONIO
  Thesis abstract: This dissertation primarily investigates the use of Digital Twin (DT) technology to enhance construction management in concrete buildings, offering an end-to-end analysis of related workflows, tools, and frameworks. It contributes to the academic field by addressing identified research challenges and supporting the advancement of DT technology both theoretically and practically.The dissertation explores the foundational concepts underpinning DTs, providing a review of their conceptual development and evolution in recent years within the construction industry. It identifies and elaborates on key technological enablers crucial to this research: i) OpenBIM (IFC standard), ii) Computational Design, iii) Internet of Things (IoT), iv) knowledge graph databases, and v) human-twin interfaces (DT platforms). Particular attention is given to the current state and challenges of integrating structural analysis into DT frameworks. A real-world case study anchors the research: the construction of a concrete office building. Through this empirical approach, six twinning information pipelines were developed, aiming to establish data flow from construction site measurements to actionable insights. These pipelines were crucial for identifying four research challenges addressed in this research: A) How can multi-layered information related to concrete construction be generated, prepared, and streamlined while ensuring accuracy and interoperability for DTs? B) What roles and workflows should stakeholders adopt to enable coordinated DT implementation? C) How can structural analysis be effectively integrated into DT systems in a scalable and interoperable manner? D) What kind of system architecture can unify diverse data layers and information pipelines to support right-time, data-driven decision-making?Research challenge A is addressed through a software development: MatchFEM. Conceived as a plugin within a computational design tool, MatchFEM streamlines the often fragmented processes of 4D IFC-BIM modeling, IoT data integration, and structural analysis by unifying them within a single parametric environment. The plugin follows a visual programming paradigm, thus simplifying the generation and preparation of DT data.The second research challenge encompasses a general mind map and two complementary workflows. They delineate the essential job roles involved in the creation and operation of DTs during building construction. Moreover, a conceptual framework for the emerging role of the DT Manager is proposed, highlighting their importance in coordinating and overseeing all DT-related activities.Furthermore, to bridge the gap between structural analysis and DT ecosystems, research challenge C, two novel data models are introduced: O-SAM (Open Structural Analysis Models), a JSON schema for encoding and transferring structural simulation data via web-based platforms, and SSO (Structural Simulation Ontology), an ontology designed to represent O-SAM data as a graph, enabling its integration within knowledge graph-based DTs.To knit all these proposals, a comprehensive DT system connected to real-time structural simulations during concrete construction is presented. This system addresses research challenge D and consolidates the knowledge, tools, and frameworks developed throughout the dissertation. The implementation of knowledge graphs as a central linking framework is emphasized, alongside the development of a human-twin interface that delivers Performance Indicators to Construction Managers. The system is validated through a prototype implementation, which incorporates an exemplary construction management workflow: the Maturity Method for concrete slabs.The dissertation concludes by reflecting on findings and contributions to the field, discussing limitations encountered, and outlining avenues for future research, envisioning DTs as data-driven assistants that enhance productivity and sustainability in the construction sector.
  

DOCTORAL DEGREE IN EARTHQUAKE ENGINEERING AND STRUCTURAL DYNAMICS

• WANG, FUMIN: A Simplified Approach for the Seismic Analysis of Compliant Soil Slopes
  
  Author: WANG, FUMIN
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN EARTHQUAKE ENGINEERING AND STRUCTURAL DYNAMICS
  Department: Department of Civil and Environmental Engineering (DECA)
  Mode: Normal
  Deposit date: 07/11/2025
  Deposit END date: 20/11/2025
  Thesis director:
  Thesis abstract: Over the past few decades, post-earthquake investigations have revealed extensive damage to both natural and engineered slopes, including landslides, slope collapses, and other forms of instability. These observations highlight the urgent need to improve our understanding of slope behavior under seismic loading and to refine analytical approaches for stability assessment. This thesis develops simplified yet robust methods for evaluating the seismic stability and dynamic performance of slopes, balancing physical insight with analytical efficiency.Chapter 2 reviews existing analytical frameworks —including the traditional Newmark rigid block, decoupled, and coupled analyses— used to estimate slope stability and dynamic responses under earthquake excitation. It compares their performance, identifies limitations, and outlines research needs.Chapter 3 proposes an extremely simplified and computationally efficient formulation for estimating the Factor of Safety (FS) of co-seismic compliant slopes, explicitly accounting for soil slope, ground motion direction, and groundwater effects. The method clarifies the differences between coupled and Newmark-type analyses and introduces a modified yield acceleration for compliant slopes. Validation through experimental, numerical, and field data confirms its simplicity and accuracy.Chapter 4 presents, as well, a coupled linear analysis based on a simplified two-block flexible model subjected to harmonic seismic excitation. Analytical and numerical solutions are derived and verified, and sensitivity analyses reveal the influence of ground motion direction, amplitude of input motion, damping ratio, and slope properties. Simplified upper bounds for maximum acceleration and displacement for the first cycle of rigid and compliant slopes are proposed, demonstrating the model’s predictive capability despite its simplicity.Chapter 5 examines the performance of classical discrete lumped-mass models against continuous wave-equation-based formulations under sinusoidal seismic loading without sliding. Both approaches yield consistent results for relative displacement and shear stress. It also shows that both models predict similar failure patterns, with cohesion playing a key role in shifting the critical sliding surface from the surface to deep depth. Incorporating Maxwell-type damping into the continuous model proves valid as an alternative to the traditional Kelvin-Voigt model in linear analysis.As a summary, this thesis extends and generalizes the simplified traditional methods for seismic analysis of soil slopes, including the stiffness of the sliding material. The proposed methodologies, validated through published experiments and real cases, provide an efficient and reliable framework for seismic slope stability assessment and offer valuable guidance for future research and engineering practice.
  

DOCTORAL DEGREE IN SIGNAL THEORY AND COMMUNICATIONS

• IRAWAN, AMIR MUSTOFA: Explainable Artificial Intelligence Applied to Geoscience and Remote Sensing: Development and Application to Wild Fire Forecasting Related to Climate Change
  
  Author: IRAWAN, AMIR MUSTOFA
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN SIGNAL THEORY AND COMMUNICATIONS
  Department: Department of Signal Theory and Communications (TSC)
  Mode: Normal
  Deposit date: 17/11/2025
  Deposit END date: 28/11/2025
  Thesis director: VALL-LLOSSERA FERRAN, MERCEDES MAGDALENA | LOPEZ MARTINEZ, CARLOS
  Thesis abstract: this thesis presents a progressive exploration of wildfire prediction by integrating process-based understanding with machine learning and causal inference frameworks. Chapter 3 focuses on variable importance and sensitivity by applying perturbation-based interventions, altering key drivers such as vapour pressure deficit (VPD), soil moisture (SM), and jet stream metrics by up to ±25% to simulate intensified environmental conditions and assess their impact on burned area. In contrast, Chapter 4 employs formal causal inference through do-calculus, enabling targeted counterfactual analysis within a structural causal model (SCM). Unlike the continuous perturbation-based interventions in Chapter 3, the intervention scenarios here are implemented by bootstrapping input variables and setting them to the 25th, 50th, 75th, and 100th percentiles. This allows the model to simulate the impact of each variable across a range of conditions, from typical to extreme (worst-case), and to quantify both direct and indirect effects on burned area, particularly for key drivers such as ∆Z500 and v300. Chapter 5 extends the causal reasoning to a global scale by using PCMCI-derived graphs as structural priors within a deep learning framework. It introduces regime-specific directed acyclic graphs (DAGs) generated through spatial clustering using the DBSCAN algorithm, enabling the identification of region-specific land–atmosphere interactions. These causal graphs are then embedded into Graph Attention Networks (GATs), allowing the model to learn weighted connections informed by causal structure, thereby enhancing both predictive performance and physical interpretability. Finally, Chapter 6 synthesizes these advances by embedding causal graphs within a GAT to simulate complex, multiscale interventions. It incorporates explicit counterfactual scenarios simulating intensified El Niño (via doubled negative SOI) and jet stream ridging (via increased positive ∆Z500, v300, and jet core), revealing spatially distinct fire responses. The use of different intervention strategies across chapters reflects the evolving methodological focus, from assessing input sensitivity (Chapter 3), to inferring causal mechanisms (Chapter 4), validating causal structures across regions (Chapter 5), and finally quantifying scenario-based outcomes (Chapter 6). Building on this foundation, Chapter 6 introduces a causal GAT capable of predicting global burned area by integrating physically grounded causal graphs derived from PCMCI. This approach enables the model to follow meaningful land–atmosphere interactions, improving interpretability and aligning predictions with known physical processes. The results show that the causal GAT outperforms models using fully connected graphs. Excessive or non-informative edges in fully connected structures can lead to over-smoothing, a common issue in Graph Neural Networks, where repeated message passing across redundant links blurs key distinctions among node representations. This can obscure critical predictive features and degrade overall model accuracy. By pruning spurious or weakly informative connections, the causal GAT preserves sharper, more meaningful node embeddings and avoids the performance loss typically associated with over-parameterized graph structures. Collectively, these advances underscore that correlation-based models fail to capture the complex, non-linear interactions among ignition sources, vegetation dynamics, and climate feedbacks. They advocate for a shift toward process-based and machine learning models that can better represent the multifaceted mechanisms governing wildfire regimes in a warming world.