Doctoral School

Reading date: 06/07/2026

• 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: 06/07/2026
  Reading time: 11:00
  Reading place: Sala de Juntes, UPC planta 1, B6
  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.
  

• GONZÁLEZ MOTOS, SERGIO: Long-term dynamics of marine microbiomes: From communities to populations
  
  Author: GONZÁLEZ MOTOS, SERGIO
  Programme: DOCTORAL DEGREE IN MARINE SCIENCES
  Department: Department of Civil and Environmental Engineering (DECA)
  Mode: Normal
  Deposit date: 05/06/2026
  Reading date: 06/07/2026
  Reading time: 10:00
  Reading place: Sala de Actos del Instituto de Ciencias del Mar (ICM) - CSICPg. Marítim de la Barceloneta, 37, Ciutat Vella, 08003 Barcelona
  Thesis director: LOGARES HAURIE, RAMIRO ERNESTO
  Thesis abstract: The surface ocean harbours a vast diversity of microorganisms, including prokaryotes (bacteria and archaea), which are key drivers of global biogeochemical cycles. Understanding how these microorganisms respond to environmental variability requires linking patterns across different biological scales, from community-level to population and strain dynamics, to the molecular consequences of adaptive mutations. Chapter 1 investigates two connected coastal microbiomes in the northwestern Mediterranean Sea: the Blanes Bay Microbial Observatory (BBMO) and the Banyuls Bay Microbial Observatory (SOLA). Using seven years of matching metagenomic data, we quantified seasonal rhythmicity and cross-site synchrony across metabolic functions, genes (ORFs), and taxa (OTUs). We found consistent seasonal rhythmicity within each microbiome across organizational levels, yet synchrony between microbiomes remained low. Focusing on 45 key biogeochemical functions, several remained highly rhythmic and synchronous even when their dominant contributing genes showed weak rhythmicity. These results suggest that functional redundancy and complementary dynamics at lower organizational levels can generate coherent functional dynamics through emergent self-organization.Moving beyond community-level patterns, Chapter 2 extended the analysis to distant ocean basins by comparing genome-resolved populations from BBMO (15 years) and MiCRO, California (10 years). We reconstructed 1,535 MAGs in BBMO and 1,068 MAGs plus 187 SAGs in MiCRO. Although the two sites are far apart, species-level overlap was considerable, with 615 genomes with intraspecific (>95% genome similarity) representatives occurring at both sites, whereas strain-level overlap was much lower, with only 17 BBMO genomes matching 20 MiCRO genomes (>99% genome similarity), revealing strong biogeographic structuring at the strain level. Genomes detected in both locations at higher similarity were more strongly associated with temperate coastal environments and tended to be slightly larger, suggesting niche adaptation to coastal conditions. Seasonal genomic variation was widespread at both sites (~70% genomes), while in MiCRO a Prochlorococcus population also showed ENSO-related increases in pN/pS and abundance during strong El Niño years, indicating that long-term climate oscillations can shape marine microbial populations beyond seasonality.Chapter 3 examined adaptation at the gene and protein levels by identifying positive selection across 1,505 MAGs from BBMO and SOLA. Positively selected genes represented less than 0.01% of the metagenomic repertoire, but several were found in both sites, pointing to non-random selection on common functions. These genes were linked to energy metabolism, stress response, regulation, translation/RNA metabolism, cell surface structures, protein turnover and motility, biofilm and environmental interactions, and exhibited contrasting seasonal patterns consistent with heterogeneous selection across cold and warm periods. Of the positively selected genes shared between BBMO and SOLA, the 70.6% also exhibited elevated pN/pS values in at least a subset of Tara samples, suggesting that many of these genes are subject to positive selection at a broader geographic scale. One key example, the RNA pseudouridine synthase (RluA), showed recurrent substitutions linked by structural modeling to increased local protein flexibility, a plausible mechanism for maintaining activity under cold conditions.
  

• MOMIN, SAMAR: An AI-Augmented Scalable Framework for High-Resolution Exposure and Seismic Risk Assessment
  
  Author: MOMIN, SAMAR
  Programme: DOCTORAL DEGREE IN EARTHQUAKE ENGINEERING AND STRUCTURAL DYNAMICS
  Department: Department of Civil and Environmental Engineering (DECA)
  Mode: Normal
  Deposit date: 13/05/2026
  Reading date: 06/07/2026
  Reading time: 16:00
  Reading place: ETSECCPB.UPC, Campus NordBuilding C1. Classroom: 002C/Jordi Girona, 1-308034 Barcelona
  Thesis director: CARREÑO TIBADUIZA, MARTHA LILIANA
  Thesis abstract: Large-scale high-resolution probabilistic seismic and multi-hazard risk assessments depend critically on how exposure is represented and how appropriately vulnerability functions are assigned to heterogeneous building stocks. While hazard modelling has achieved a high degree of harmonisation and vulnerability modelling has achieved a high level of computational maturity, the exposure–vulnerability interface remains manual, typology-driven and difficult to scale consistently. In Barcelona and across Catalonia, vulnerability functions derived from single buildings or limited samples have frequently been extrapolated across portfolios using coarse classifications. When hazard, exposure and damage-to-loss models are held constant, this practice produces substantial spreads in portfolio loss metrics. It first quantifies that variability through a controlled vulnerability-only sensitivity analysis for Barcelona. Using a consistent probabilistic seismic hazard backbone and unified damage-to-loss mapping, published vulnerability-function sets are interchanged while all other components remain fixed. Results show that Average Annual Loss (AAL) varies from approximately €25.6 million to €348.0 million depending solely on the selected vulnerability dataset, demonstrating order-of-magnitude sensitivity attributable to function assignment. A focused pilot in the Eixample district confirms that detailed, manual reassignment (e.g., corner versus centre block positions) improves local realism but does not scale to city- or regional-level portfolios.To address this structural limitation, the second step moves from ad-hoc assignment toward repeatable and transparent exposure. The thesis develops a geometry-driven exposure structuring workflow through Building Footprint Analysis, Classification and Grouping (BFA/BFC/BFG). High-resolution cadastral building footprints from Catalonia are used to extract dimensions and are transformed into interpretable, reproducible plan-shape classes using previously computed shape metrics to define sides-based classes. Results from comparative analysis of several shape metrics show that simple compactness measures agree on regular rectangles and squares but blur important differences for elongated, irregular and opening-based plans. This reveals redundancy among scalar compactness indices and limited discriminatory power for articulated and opening-based geometries, highlighting the need for an alternative capable of preserving fine-grained distinctions relevant to structural behaviour.The Building Footprint Analysis, Recognition and Classification framework (BFARC-YOLO) builds on this basis to enable scalable AI-augmented exposure classification. Approximately 1.8 million building footprints are rendered as oriented silhouettes and progressively trained using ~0.34% labelled samples. The final model achieves mean average precision (mAP@0.5:0.95) ≈ 0.96 across 44 fine-grained classes. GPU optimisation reduces regional inference time from over 490 hours (CPU baseline) to ~72 hours, demonstrating operational feasibility. Interoperability with the GEM taxonomy enables direct integration into established risk platforms and a lightweight web interface supports auditable batch inference.Rather than developing novel fragility functions, this thesis offers modelling infrastructure. It creates a replicable bridge from cadastral geometry to vulnerability-aligned exposure classification, minimising subjectivity in vulnerability assignment and allowing for scalable, transparent portfolio risk modelling. By reframing exposure as a structured, learning-enabled modelling layer, the study lays the groundwork for future fragility calibration, and next-generation multi-hazard risk assessments for dense and heterogeneous cities, regions, countries and the global scale.
  

• PUIG DE DOU, IGNACIO: Statistical Activity Tracking: Bayesian Hierarchical Models for Monitoring Customers and Industrial Equipment
  
  Author: PUIG DE DOU, IGNACIO
  Programme: DOCTORAL DEGREE IN STATISTICS AND OPERATIONS RESEARCH
  Department: Department of Statistics and Operations Research (EIO)
  Mode: Article-based thesis
  Deposit date: 12/06/2026
  Reading date: 06/07/2026
  Reading time: 11:00
  Reading place: Sala d'Actes de la Facultat de Matemàtiques i Estadística
  Thesis director: PUIG ORIOL, XAVIER
  Committee:
       PRESIDENT: FERRER RIQUELME, ALBERTO JOSÉ
       SECRETARI: TORT-MARTORELL LLABRES, FRANCISCO JAVIER
       VOCAL: TSIAMYRTZIS, PANAGIOTIS
  Thesis abstract: Modern business and industrial environments increasingly rely on continuous monitoring of customers and equipment through operational data streams. Detecting meaningful behavioral changes, such as customer disengagement or early signs of equipment malfunction, is critical for timely intervention. Classical Statistical Process Control (SPC) methods rely on assumptions that are often violated in practice, including homogeneous populations, regular exposure and a clear separation between Phase I and Phase II.This thesis develops Bayesian modeling and methodological solutions for tracking heterogeneous entities, such as customers or remotely operated industrial machines, under these practical constraints. Irregular activity, missing data, evolving operating conditions and the absence of a clean Phase I period are treated as defining features of modern monitoring problems rather than as limitations to be circumvented.The first part of the thesis focuses on customer-activity modeling through a fully Bayesian implementation of the Pareto/NBD model framework. A Bayesian clustering strategy is introduced to group customers with similar purchasing behavior, improving parameter estimation and supporting customer-level revenue forecasting.The second part addresses industrial equipment monitoring using error-count data. Two complementary approaches are proposed. First a mixture-model-based SPC framework for heterogeneous populations when in-control data are not clearly identifiable. Second a fully Bayesian hierarchical model that incorporates operating conditions, stabilizes low-exposure estimates, captures machine-specific behavior and automatically identifies out-of-control observations.Overall, Bayesian inference provides a flexible and principled framework for hierarchical modeling, probabilistic anomaly detection, information sharing across entities and sequential updating as new data become available. The proposed methods extend classical SPC ideas to contemporary business and industrial monitoring settings, offering statistically sound and operationally relevant solutions.
  

Reading date: 07/07/2026

• HERNANDEZ RUIZ, MARÍA: Cavity-Enhanced, Optically Pumped Magnetometry and its Miniaturization
  
  Author: HERNANDEZ RUIZ, MARÍA
  Programme: DOCTORAL DEGREE IN PHOTONICS
  Department: Institute of Photonic Sciences (ICFO)
  Mode: Normal
  Deposit date: 26/05/2026
  Reading date: 07/07/2026
  Reading time: 15:00
  Reading place: ICFO Auditorium
  Thesis director: MITCHELL, MORGAN | LUCIVERO, VITO GIOVANNI
  Thesis abstract: This thesis presents the development and experimental implementation towards a miniaturized cavity-enhanced atomic magnetometer designed for high sensitivity and high spatial resolution. The proposed technique relies on a cavity-based optical readout of atomic polarization using the Pound Drever Hall technique.The magnetometer was experimentally realized and characterized, including a design compatible with biological measurements. The magnetometer can operate with one endcap open in a weak shielding environment (only only layer of mu-metal.) The system was applied to the study of magnetotactic bacteria, whose magnetic response arises from the alignment and relaxation dynamics of intracellular magnetosome chains. Time resolved measurements of the magnetic signal were analyzed assuming exponential relaxation. Furthermore, the model was expanded using a stochastic model to detect deviations from exponential behavior associated with sample evaporation.The achieved sensitivity and stability of the magnetometer enabled the reliable detection of magnetic signals on the nanotesla scale, with a sensing characteristic length Lat=1.5 mm. These results demonstrate the potential of cavity-enhanced atomic magnetometry for the investigation of biological magnetic systems. This work establishes a versatile platform for high-sensitivity magnetic measurements and opens promising perspectives for future applications in biophysics and magnetic microscopy.