Doctoral School

Reading date: 08/06/2026

• BAHRAMIAN, LINDA: Numerical assessment of parcel modeling and inertial particle separator efficiency in polydisperse two-phase flows
  
  Author: BAHRAMIAN, LINDA
  Programme: DOCTORAL DEGREE IN THERMAL ENGINEERING
  Department: Department of Heat Engines (MMT)
  Mode: Normal
  Deposit date: 11/05/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: OLIET CASASAYAS, CARLES | PEREZ SEGARRA, CARLOS DAVID
  Thesis abstract: This thesis addresses fundamental and applied aspects of disperse two-phase flow modeling within the Eulerian–Lagrangian approach, with particular emphasis on the development and assessment of parcel modeling and the conservation-consistent two-way coupling approach, progressing toward application of an Inertial Particle Separator (IPS).First, the conservation properties of the two-way coupling formulation are analyzed to ensure consistent momentum exchange between the carrier and dispersed phases. The numerical implementation preserves global conservation principles and physically consistent kinetic energy evolution.The numerical method is first assessed in Direct Numerical Simulation to provide a reference solution and is subsequently extended to Large Eddy Simulation (LES) to address more realistic, application-oriented flow conditions. Building upon this foundation, a novel hybrid parcel modeling strategy is developed by combiningthe Number Fixed Model and the Volume Fixed Model for a particle distribution. The proposed approach is then validated against benchmark cases and offers an effective compromise between computational cost and predictive accuracy.Subsequently, a comprehensive numerical investigation of an IPS device is performed using LES, Improved Delayed Detached Eddy Simulation, and Reynolds-Averaged Navier–Stokes (RANS) turbulence models. The influence of turbulence resolution, flow split ratio, and Reynolds number on separation efficiency is analyzed for different particle sizes. The results indicate that turbulence modeling affects drag-dominated particles, while segregation of larger particles is mainly driven by inertial effects. This study presents a comparative assessment of turbulence modeling strategies and their impact on IPS performance prediction.The numerical framework is then extended to icing conditions. Water droplet impingement and ice accretion are first validated using a canonical cylinder benchmark. The methodology is then applied to the IPS configuration, where ice growth alters the internal geometry. Following ice accretion, solid particles are injected into the modified geometry to evaluate the separation efficiency using Lagrangian tracking, highlighting a reduction in efficiency for inertia-dominated particles as a result of scavenge blockage and modified wall interactions.In this context, a key original contribution of this study is the development and implementation of a dedicated particle–ice wall collision model. To address the lack of suitable models for solid particles impact on ice-covered surfaces in IPS, a restitution-based formulation was developed through a structured review and adaptation of existing collision models.This approach ensures a physically consistent prediction of particle rebound behavior on ice-covered surfaces.Overall, this thesis advances the predictive capability of Computational Fluid Dynamics tools for disperse two-phase flows by enhancing the implementation of conservation-consistent two-way coupling, proposing a physics-guided hybrid parcel model, quantifying the sensitivity of turbulence modeling in IPS efficiency, and introducing a novel particle–wall collision model under adverse icing conditions. The results contribute to both the methodological development of Eulerian–Lagrangian modeling and the reliable simulation of aeronautical particle separation systems.
  

• CAZZARO, FRANCESCO: Advancing Text-to-Query Semantic Parsing Systems
  
  Author: CAZZARO, FRANCESCO
  Programme: DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE
  Department: Department of Computer Science (CS)
  Mode: Normal
  Deposit date: 07/05/2026
  Reading date: 08/06/2026
  Reading time: 11:00
  Reading place: sala de juntes de la FIB, edifici B6 - Campus Nord UPC (Barcelona)
  Thesis director: QUATTONI, ARIADNA JULIETA
  Thesis abstract: Executable Semantic Parsing is the task of mapping a natural language sentence into a formal meaning representation that can be executed over a knowledge base to retrieve information. This task presents several challenges. Natural language variability makes the translation process inherently difficult and, at the same time, meaning representations are highly compositional structures built from elementary units. As a result, semantic parsers must exhibit strong compositional generalization, a capability that remains challenging for current models. Moreover, annotated data for semantic parsing is scarce, and its collection is labor intensive and costly, making it difficult to train models.In this thesis, our aim is to improve and advance semantic parsing systems. We introduce a pipeline that decouples the semantic parsing process into two steps, a translation stage and a reordering stage, which enhances the compositional abilities of the parser. We also propose a data generation method that recombines existing annotated pairs in novel ways. This approach improves the generalization capabilities of semantic parsers while alleviating the data scarcity problem.Furthermore, to address the challenge of limited annotated data, we design a novel approach to automatically generate data pairs from a given knowledge graph without the need of human intervention. This generated data can then be used to train a semantic parsing system specifically designed for that particular knowledge graph. This thesis is among the first to explore semantic parsing for property graphs, where we not only introduce the data generation method but also provide an annotated benchmark for evaluating parsing performance.
  

• DE LIMAS SANTANA, ALEXANDRE: Beyond 512-bit vectors: Optimized and performance portable AI operators on vector-length-agnostic architectures
  
  Author: DE LIMAS SANTANA, ALEXANDRE
  Programme: DOCTORAL DEGREE IN COMPUTER ARCHITECTURE
  Department: Department of Computer Architecture (DAC)
  Mode: Normal
  Deposit date: 11/05/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: CASAS GUIX, MARC | ARMEJACH SANOSA, ADRIÀ
  Thesis abstract: This thesis addresses the challenges of producing high-performance, portable code for arithmetic-intensive operations in Deep Neural Network (DNN) workloads, such as convolutions and matrix multiplications.It targets emerging vector and matrix architectures exposing software interfaces to program accelerators via Instruction Set Architecture (ISA) extensions.It proposes techniques to handle three main challenges related to accelerating DNN workloads on contemporary data-parallel CPU architectures: i) extrapolating software optimization techniques designed primarily for high-end general-purpose vector processors to also accomodate edge devices and long vector accelerators hardware, ii) providing performance portability for open standard vector ISA ecosystems, like RISC-V, characterized by their unprecedent micro-architectural discrepancy among implementations, and iii) designing matrix multiplication ISA extensions that uphold the core principles of emerging vector architecture concerning implementation flexibility, such as vector length agnosticism.This thesis presents empirical evidence that existing software optimization techniques for generating high-performance implementations of DNN operators for modern vector processors are biased toward high-end general-purpose systems with 512-bit vectors (e.g., Intel Cascade Lake, Fugaku AF64X).The techniques presented in the thesis support the idea of hardware/software co-design and the need to consider microarchitectural features, such as vector length, non-conventional memory subsystems, and the presence/lack of out-of-order pipelines, when generating code for vector and matrix processors.Specifically, this thesis employs runtime specialization to adapt established DNN algorithms to a broader range of processor designs, equipping applications with flexible code generators that dynamically optimize the algorithms for the specific combination of the platform's microarchitectural features and the operation's hyperparameters. The thesis makes three key contributions.First, it provides the first performance analysis of convolution workloads on a 16,384-bit vector processor, identifies cache conflict misses in existing techniques, and proposes software corrections that yield up to 1.83x speedups over prior approaches.Second, it introduces a dynamic matrix multiplication and convolution code generator that adapts algorithmic optimization techniques, such as register unrolling, to the vector width of the target platform.A performance analysis of three commercially available RISC-V systems running computer vision workloads shows geometric speedups ranging from 1.43x to 3.58x compared to state-of-the-art libraries.Third, it proposes a Matrix Tile Extension (MTE) to supplement vector-length-agnostic ISAs with geometry-agnostic matrix multiplication and memory operations.Two microarchitectures supporting this extension are described: a lean extension of an 8192-bit vector processor and a systolic-array-based design, both of which utilize vector registers for matrix storage and demonstrate 1.35x geometric mean speedups on GEMM and convolution workloads.
  

• GUO, ZHIMING: Study on HTPB propellant passivation, mixing, casting and curing processes by experiment and simulation
  
  Author: GUO, ZHIMING
  Programme: DOCTORAL DEGREE IN STRUCTURAL ANALYSIS
  Department: Department of Civil and Environmental Engineering (DECA)
  Mode: Article-based thesis
  Deposit date: 11/05/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: ROSSI BERNECOLI, RICCARDO | FU, XIAOLONG
  Thesis abstract: This study investigates the key physical issues involved in the four propellant production steps (passivation, kneading, casting, and curing) through a combination of experimental and numerical simulations.In this experimental study using HTPB propellant ingredients as raw materials, we first investigate the passivation and dehydration (similar to reduced pressure micro-boiling) of this raw material (N-butylnitroxyethylnitramine (BuNENA)). Next, we add other materials (including liquids and granules) to the passivated raw material and mix them in a vertical kneader. The resulting propellant slurry is then cast into a specific mold. Finally, we investigated the solidification of the propellant samples formed in the mold.In this numerical simulation study, First, the passivation process (bubble generation and movement) of the BuNENA material was studied through numerical simulation. Next, the mixing process of the propellant in a vertical kneader was investigated. The uniformity and flow characteristics of the HTPB propellant material were studied under stirring conditions.Finally, the slurry casting process was simulated, and finally, the curing of the cast propellant model was simulated.This paper contains the following research contents:(1) A passivation experimental device was established and experiments were carried out using a principle similar to reduced pressure micro-boiling; in the fluid dynamics simulation model, the Lagrangian framework was applied to track the formation and movement of bubbles, and the bubbles themselves were modeled as rigid spheres subjected to buoyancy and viscous forces. The Euler framework based on variational multiscale (VMS) was used to simulate the fluid around the bubbles. The bubble movement was analyzed. By combining experimental and simulation methods, the passivation process of BuNENA was analyzed in detail, which is of substantial significance in the field of passivation of composite solid propellants.(2) A computational fluid dynamics (CFD) method was used to establish a digital simulation model of the mixing process of the vertical kneader. Changes in various flow field related characteristics of the vertical kneader were analyzed. The mixing performance of the propellant slurry in the kneader was studied by establishing a mixing uniformity index analysis method. The accuracy of the simulation results was verified by real kneading experiments, SEM-EDS and density experiments. (3) The vacuum casting process was optimized by combining experiments and numerical simulations. First, the shear thinning behavior was revealed through rheological tests, and the Herschel-Bulkley model parameters confirmed non-Newtonian fluid characteristics. The variational multi-scale finite element method was used to simulate and analyze the vacuum casting process of HTPB propellant slurry. Second, the flow rate and impact force of droplets under different vacuum pressures were studied by combining real-time image recognition with machine vision and Kalman filtering. (4) Finally, the curing reaction kinetic model of HTPB propellant was studied by the non-isothermal DSC method. The distribution and evolution of the internal temperature and temperature degree of the propellant during the molding process were analyzed using a thermochemical model. The temperature gradient and curing time variation of the propellant curing process were explored by the thermocouple-integrated method.
  

• HERNÀNDEZ CARNERERO, ÀLVAR: Generalizable Time Series Classification Models for Complex Real-World Applications
  
  Author: HERNÀNDEZ CARNERERO, ÀLVAR
  Programme: DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE
  Department: Department of Computer Science (CS)
  Mode: Normal
  Deposit date: 23/04/2026
  Reading date: 08/06/2026
  Reading time: 12:00
  Reading place: Sala d'actes de la FIB - Edifici B6 planta 0, Campus Nord
  Thesis director: SANCHEZ MARRE, MIQUEL | VAZQUEZ SALCEDA, JAVIER
  Thesis abstract: This thesis develops a comprehensive framework for Time Series Classification (TSC) focused on real-world applications characterized by limited data availability, heterogeneous datasets, and complex temporal structure, which concludes with a generalizable and adaptive modelling approach. This research conducts a comprehensive survey of state-of-the-art methods, identifying their strengths and limitations, with a focus on specific domains such as antimicrobial prediction, astronomy, and physiological monitoring. Building on these insights, preprocessing pipelines are developed for the considered time series, including new temporally informed feature generation grounded in domain knowledge, robust feature importance estimation, and wrapper-based feature selection. Methodologically, windowing strategies are engineered to mitigate data scarcity and temporal distribution shifts, a hybrid modular architecture is proposed to integrate mechanisms suited to the specific needs of the data, and an adaptation of additive attention is formulated. The methodology employs ensembles to enhance predictive stability and evaluates model interpretability by analyzing how the trained models prioritize features and by visualizing attention patterns across the sequence to reveal temporal relevance patterns.A central contribution of the thesis is CLAIM (Content-modulated Low-rank Adaptive Isotropic Mixer), a novel architecture for TSC that synthesizes insights from our research and from recent advances in sequence modelling. CLAIM combines content-modulated dynamic mixing, adaptive inductive biases, and a low-rank isotropic block design, resulting in an efficient and interpretable model. Experiments across heterogeneous datasets (including intensive care unit patient trajectories, exoplanet transit signals, and electrocardiogram waveforms) show that CLAIM frequently achieves the best performance in small-data regimes compared to temporal transformer and Multilayer Perceptron Mixer (MLP-Mixer) baselines. CLAIM also offers practical advantages in interpretability, due to its dynamic weights and explicit inductive-bias components. In addition, it demonstrates strong computational efficiency, particularly at inference time. The model modulates structural priors such as locality, causality, and location invariance in a data-driven manner, aligning with effective contemporary design principles for temporal modelling. Furthermore, it adapts its focus across domains, as reflected in the dynamic weighting of temporal components, which corresponds with established domain knowledge.The thesis contributes a unified perspective on real-world TSC, pairing domain-guided pipelines with a generalizable modelling approach. The resulting CLAIM architecture offers an interpretable, efficient, and adaptive solution that advances the state of the art in small-data TSC.