Doctoral School

Reading date: 26/11/2025

• JAMES, CHRISTOPHER WILLIAM VINCENT: Biomaterials for Cardiac Regeneration
  
  Author: JAMES, CHRISTOPHER WILLIAM VINCENT
  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 BIOMEDICAL ENGINEERING
  Department: Department of Materials Science and Engineering (CEM)
  Mode: Normal
  Deposit date: 01/10/2025
  Reading date: 26/11/2025
  Reading time: 16:00
  Reading place: Sala d'Actes, Edifici Vèrtex, Campus Diagonal Nord, Vèrtex (VX), Plaça d'Eusebi Güell, 6, 08034 Barcelona
  Thesis director: ENGEL LOPEZ, ELISABET
  Thesis abstract: Cardiovascular diseases are the leading cause of death globally, with heart transplantation being the most effective treatment following injury due to the heart’s limited regenerative capacity. In situ tissue engineering has emerged as a promising approach to activate endogenous cardiac repair. This thesis focuses on the design, development, and characterization of injectable biomaterials for in situ cardiac regeneration.Hydrogel scaffolds that physically support damaged tissue, provide inherent bioactivity, or deliver bioactive agents are particularly promising. We first isolated and characterized porcine cardiac extracellular matrix (cECM) hydrogels, demonstrating their compatibility for supporting the growth of cardiac-associated cells. To enhance the viscoelastic properties of cECM without compromising biocompatibility, alginate—a hydrogel currently in clinical trials—was incorporated, resulting in improved mechanical properties.Lactate, traditionally seen as a metabolic by-product, has gained attention for its role in promoting angiogenesis, cardiomyocyte proliferation, and reducing fibrosis. Based on this evidence, we developed a lactate-release scaffold by embedding FDA-approved PLGA nanoparticles into the cECM-alginate matrix, optimizing nanoparticle size and degradation to achieve sustained lactate release.The regenerative potential of stem cell-derived secretomes, comprising bioactive molecules such as growth factors and extracellular vesicles, was also explored. Bone marrow-derived stromal cell (BMSC) secretomes were evaluated for their cardioprotective effects on human cardiac fibroblasts. A novel culture method showed superior outcomes, and the resulting secretome was incorporated into the scaffold either directly or via PLGA nanoparticles for sustained delivery.In conclusion, this work presents several novel injectable biomaterials that show potential for in situ cardiac regeneration through enhanced mechanical support, bioactivity, and sustained delivery of regenerative agents such as lactate and BMSC-derived secretomes. These findings warrant further investigation to optimize the therapeutic efficacy of these platforms.
  

• RIU I VICENTE, JORDI: Scaling Quantum Optimization Algorithms: Advancing Techniques for Handling Industrial-Level Workloads with Artificial Intelligence
  
  Author: RIU I VICENTE, JORDI
  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: 01/10/2025
  Reading date: 26/11/2025
  Reading time: 11:00
  Reading place: Sala d'Actes, Edifici Vèrtex, Plaça d'Eusebi Güell 6, Barcelona
  Thesis director: GARCIA SAEZ, ARTURO | MONRAS BLASI, ALEXANDRE
  Thesis abstract: This thesis integrates traditional optimization methods and artificial intelligence techniques with near‑term quantum algorithms to accelerate the adoption of hybrid protocols for large‑scale, time‑sensitive decision problems. These techniques can manage millions of variables, but they are very expensive computationally. Conversely, quantum approaches such as quantum annealing and QAOA promise more efficient exploration of complex solution landscapes, yet they are hampered by severe hardware limitations and steep training challenges like barren plateaus. By combining these paradigms, the work offloads subproblems suited to classical resources and reserves quantum circuits for the most demanding kernels.We first present the formulation of a relevant use case of a medical‑drone logistics network modeled across Catalonia’s mountainous terrain. An initial discrete‑time MILP with millions of binary variables captures the relevant properties but exceeds the reach of current quantum hardware. A continuous‑time reformulation later reduces variable count. Resource estimates confirm that hardware scaling alone cannot close the gap, underscoring the need for new algorithmic strategies.To address NISQ‑era circuit depth, the thesis introduces RL‑ZX, a reinforcement‑learning–driven quantum compiler built on the ZX‑Calculus framework. Quantum circuits translate into feature‑annotated ZX diagrams, which a Graph Attention Network encodes. An agent trained with proximal policy optimization then applies ZX rewrite rules to minimize two‑qubit gates under device‑specific cost metrics. When evaluated on large, unseen circuits, RL‑ZX outperforms leading heuristics almost universally, yielding shallower, higher‑fidelity circuits.Next, we automate the process of including hard constraints in QUBO formulations. We design a Graph Neural Network and train it to predict penalty magnitudes in a single inference pass, treating variables and constraints as nodes and edges annotated with energy‑difference features. Integrated end‑to‑end with a differentiable QAOA simulator, this approach uses ground‑state sampling probability as its training signal. Experiments on random instances of the knapsack problem, TSP, and assignment tasks demonstrate large improvements in solution probability compared to analytic bounds.The third topic under study is the choice of starting state for VQAs. By measuring coherence via relative entropy, we show that higher‑coherence initial states produce better approximation ratios in fixed‑depth QAOA on small Max‑Cut instances. A tensor‑network imaginary‑time evolution protocol generates Matrix‑Product States approximating pure Gibbs states, that are later mapped to shallow quantum circuits. This initialization strategy suggests a clear path toward improved performance as quantum hardware matures.Finally, two “constraint-satisfying” ansätze are developed to prepare valid solutions directly, eliminating penalty terms. The first builds assignments incrementally using multi‑controlled rotations, while the second creates superpositions of partial assignments before deterministically completing them with ancilla qubits.By uniting these advances—hybrid classical/quantum decomposition, RL‑driven compilation, GNN‑based penalty tuning, coherence‑aware initialization, and feasible‑only state preparation—the thesis lays a scalable, modular foundation for achieving quantum advantage in real‑world optimization.
  

Reading date: 27/11/2025

• ALLKA, XHENSILDA: Enhancing Data Quality in IoT Monitoring Sensor Networks
  
  Author: ALLKA, XHENSILDA
  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: 31/10/2025
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: BARCELÓ ORDINAS, JOSE MARIA | GARCÍA VIDAL, JORGE
  Thesis abstract: In recent years, technological development and an increased number of cars among other factors, have influenced air pollution levels. This increase in levels has also increased the need to monitor them, as they are directly related to human health and the economy. To monitor air pollution, the government has deployed precise monitoring stations, which are expensive to deploy and maintain. Due to their cost, they are not widely distributed. However, since air pollution can change over short distances, the distribution of these stations can be insufficient. Recently, a solution has emerged: the use of low-cost sensors (LCSs), which provide broader coverage at a much lower cost. However, these LCSs have one drawback: the quality of the data they provide is poor.Current research in this field has employed machine learning (ML) models to calibrate these LCSs, thereby enhancing the quality of the data they provide. In an Internet of Things (IoT) monitoring network, the quality of data is closely associated with decision-making processes. This thesis focuses on enhancing the data quality provided by the LCSs from two perspectives: improving calibration performance and detecting anomalies and outliers. The objective of both of these perspectives is to ensure data accuracy and reliability.The first part of the thesis focuses on the improvement of the calibrated data provided by the LCSs and the detection of the concept drift and the update of the parameters of the current calibration model such that it adapts to the new conditions. We are enhancing the quality of the calibrated data by implementing a model pattern-based approach. Our proposed methods, Temporal Pattern Based Denoising (TPB-D) and Temporal Pattern Based Calibration (TPB-C), improve the quality of the calibrated data. Given that environmental conditions are subject to change over time, it is essential to update the parameters of the calibration model. We proposed Window-based Uncertainty Drift Detection and Recalibration (W-UDDR), a system capable of detecting the presence of concept drift (i.e., environmental changes).The second part of the thesis focuses on the reliability of the data. Sensors, regardless of their cost, are often prone to irregularities such as outliers, anomalies, or drift, which can be caused by various factors. It is critical to identify these irregularities, as the data will be incorporated into the training of the model related to other tasks. In this thesis, three distinct scenarios were examined. The first one is related to the detection of outliers in the edge. In this case, we proposed the Edge Streaming Outlier Detection (ESOD) framework. ESOD is a simple and lightweight framework that can identify outliers in the edge with a limited amount of memory. The system offers two approaches: real-time and near real-time. The near real-time approach involves minor delays in decision-making. The second approach is related to the detection of more complex irregularities, such as anomalies in a given sensor. This scenario is distinct from the first one in that it offers offline anomaly detection capabilities. We proposed spatiotemporal correlation recurrent autoencoder anomaly detection (STC-RAAD), which demonstrated satisfactory performance in detecting anomalies in a given sensor. It is worth noting that the third scenario pertains to the detection and localization of anomalies in a network of sensors. This is of particular relevance in scenarios where the identification and precise location of the source of an anomaly are crucial. We hereby propose a pattern-based attention recurrent autoencoder anomaly detection (PARAAD) method. This method is designed to detect and localize anomalies in sensors.
  

• CARRILLO CERVERA, ALEJANDRO: Efecto del área verde en la salud: parques y enfermedades cardiovasculares en Mérida (2020)
  
  Author: CARRILLO CERVERA, ALEJANDRO
  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 URBAN AND ARCHITECTURAL MANAGEMENT AND VALUATION
  Department: Department of Architectural Technology (TA)
  Mode: Normal
  Deposit date: 25/09/2025
  Reading date: 27/11/2025
  Reading time: 17:00
  Reading place: ETSAB (Esc. Téc. Sup. Arquitectura de Bcn)-Pl. Baja-Sala GradosEnlace videoconf.: https://meet.google.com/vzy-djqq-sqaInicio conexión 16:30 h
  Thesis director: ROCA CLADERA, JOSE NICASIO | ARELLANO RAMOS, BLANCA ESMARAGDA
  Thesis abstract: This doctoral study analyzes the relationship between urban green spaces and cardiovascular disease (CVD) mortality in the city of Mérida, Yucatán (Mexico), using a pioneering methodology at the urban block scale. The research combined spatial analysis, statistical techniques, and direct observation to assess the impact of green coverage, quality (NDVI), and accessibility to urban parks on public health.In Mérida, the analysis reveals that daily proximity to green areas—especially small or medium-sized parks within 300 meters—has a statistically significant effect (β = –0.001; p < 0.001). These parks exhibit the highest quality of urban greenery and the best plant health. In contrast, larger parks over 5 hectares, where vegetation quality and plant health are poor, show no significant effect (β = 0.000; p = 0.006). This suggests that it is not only the amount of permeable land in parks that matters, but also the quality of urban vegetation and accessibility in the immediate surroundings. By age group, young adults (18–24 years) had better access to parks, while children and older adults faced more barriers.People with disabilities are more vulnerable to developing CVD (β = 0.324; p = 0.010). The findings also indicate an increase in mortality associated with the use of private vehicles as a means of transportation (β = 0.261; p = 0.007). On the other hand, higher spending on food consumed at home has a protective effect (β = –0.165; p < 0.001).Although other factors may also explain changes in mortality rates, urban greenery and the quality of the built environment significantly influence population health. The regression equation was statistically significant F(5, 10036) = 0.001. The R² value was 0.202, indicating that 20.2% of the CVD mortality rate can be explained by the model, which includes variables such as accessibility to urban parks, the percentage of people with disabilities, spending on private vehicle use, and spending on food consumed at home.The mortality rate decreases by –0.001 points for every additional square meter of urban greenery in parks. Excluding medians and sidewalks, the population has access to an average of 7.17 m² of urban green space per inhabitant across all parks within the peripheral ring, and 5.79 m² per inhabitant when considering only parks of 5 hectares or more. The average CVD mortality rate is 29.78% per urban block.
  

• DELGADO GUERRERO, JUAN ANTONIO: Learning latent structures for robotic assistance in daily manipulation tasks
  
  Author: DELGADO GUERRERO, JUAN ANTONIO
  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 Robotics and Industrial Informatics (IRI)
  Mode: Normal
  Deposit date: 02/09/2025
  Reading date: 27/11/2025
  Reading time: 11:00
  Reading place: Sala d'Actes Manuel Marta Recober, planta 0, Edifici B6, Facultat d'Informática de Barcelona (FIB), Campus Nord, C/Jordi Girona, 1-3, 08034 Barcelona
  Thesis director: TORRAS GENIS, CARMEN | COLOMÉ FIGUERAS, ADRIÀ
  Thesis abstract: Robotic domestic assistance presents significant challenges due to the complexity of modeling everyday manipulation tasks, especially those involving deformable objects like cloth. Traditional approaches often struggle with high-dimensional state representations, dynamic uncertainties, and the need for safe human-robot interaction. This thesis addresses these challenges by developing novel machine learning methods based on latent variable models to enable efficient, adaptive, and safe robotic manipulation.First, we propose a Gaussian Process Latent Variable Model (GPLVM) framework combined with Bayesian Optimization (BO) to learn high-dimensional robot motion policies with minimal data. This approach reduces the parameter space dimensionality while preserving task-relevant features, achieving faster convergence than other existing model-free alternatives.Next, we extend this framework to contextual learning using Covariate GPLVM (c-GPLVM), allowing robots to adapt to environmental changes (e.g., user preferences, object positions) without retraining. Experiments in feeding and shoe-fitting tasks demonstrate improved generalization with fewer samples compared to state-of-the-art contextual policy search methods.For dynamic cloth manipulation, we introduce the Controlled Gaussian Process Dynamical Model (CGPDM), which embeds control actions into a low-dimensional latent space to predict cloth motion under robot manipulation. Evaluations in simulated and real-world bimanual cloth handling show that CGPDM accurately generalizes to unseen actions, even with limited training data.Finally, we address safety in human-robot interaction by proposing Cartesian control enhancements for redundant manipulators, including error saturation, singularity avoidance, and impedance tuning. These measures mitigate risks during physical interaction, ensuring stable and compliant robot behavior.Together, these contributions advance robotic cloth manipulation by combining data-efficient learning, context-aware adaptation, and safe control, paving the way for practical deployment in assistive and household robotics.