Doctoral School

Reading date: 15/06/2026

• VILALTA ARIAS, ARMAND: Semantic Embeddings in Deep Convolutional Neural Networks
  
  Author: VILALTA ARIAS, ARMAND
  Programme: DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE
  Department: Department of Computer Science (CS)
  Mode: Normal
  Deposit date: 26/03/2026
  Reading date: 15/06/2026
  Reading time: 11:00
  Reading place: Sala d'actes de la FIB - Edifici B6, Campus Nord Barcelona
  Thesis director: GARCÍA GASULLA, DARIO | CORTÉS GARCÍA, CLAUDIO ULISES
  Thesis abstract: One of the fundamental questions in Artificial Intelligence (AI) is how to represent knowledge effectively. The main challenge lies in devising a representation that can be exploited for intelligent computations. Since 2012, when Deep Neural Networks (DNNs) established themselves as the state of the art in Computer Vision, their use has rapidly expanded to a wide range of AI domains, including text and speech recognition, image generation, gaming, and robotics. This widespread success stems from their nature as representation learning techniques. However, DNNs require large volumes of data and considerable computational resources to learn such representations, which significantly limits the range of problems to which these models can be directly applied.In this thesis, we explore the possibility of reusing the knowledge learned by a DNN for a specific problem to solve a different one, which is commonly known as transfer learning. We focus on the representations learned by a DNN, codified in the activations of its neurons in response to an input, namely embeddings. All this work stems from a first analysis to understand where valuable information is encoded in Convolutional Neural Networks (CNNs) within the context of transfer learning. When used for characterizing every class of eleven datasets, we statistically measure the discriminative power of every feature found within a deep CNN. We seek to provide new insights into the behaviour of CNN features, particularly the ones from convolutional layers, which they had not used in previous literature. Our results confirm that low and middle level features may behave differently from high-level features, but only under certain conditions. We find that all CNN features can be used for knowledge representation purposes both by their presence or by their absence, doubling the information a single CNN feature may provide. We also study how much noise these features may include and propose a thresholding approach to discard most of it.This enables the definition of a methodology to improve the generalization capabilities of CNN representations, which we refer to as the Full-Network Embedding (FNE), which successfully integrates convolutional and fully connected features. To do so, the embedding normalizes features in the context of the problem and discretizes their values to reduce noise and regularize the embedding space. Significantly, this also reduces the computational cost of processing the resultant representations. The proposed method outperforms single layer embeddings on several image classification tasks while also being more robust to the choice of the pre-trained model used for obtaining the initial features. The performance gap in classification accuracy between thoroughly tuned solutions and the full-network embedding is also reduced, making the proposed approach a competitive solution for a large set of applications. We consider FNE more semantic as it understands CNN activations as semantic concepts. Similarly to many human languages, it considers if a feature of the input is significant in a given context and if it is because of its presence or absence.A first stem is the use of the FNE as the basis for a network representation of concepts where complex network techniques are applied. We propose the construction of a graph embedding space, introducing a methodology to transform the knowledge coded within a deep convolutional network into a topological space.The second stem, uses the same techniques in the context of multimodal embeddings representing text and images. The FNE provides a multi-scale representation of images, which results in richer characterizations while focusing on the relevant information through its semantic discretization. Results for image annotation and image retrieval tasks show a constant improvement when applied to different existing methodologies.
  

Reading date: 16/06/2026

• PAMIES SAURET, CARLES: Sillerías góticas españolas. Reconstrucción digital mediante fotogrametría de las sillerías góticas españolas: Análisis de la ubicación y funciones del Coro en las Catedrales de los siglos XV-XVI
  
  Author: PAMIES SAURET, CARLES
  Programme: DOCTORAL DEGREE IN ARCHITECTURAL, CIVIL AND URBAN HERITAGE AND REFURBISHMENT OF EXISTING BUILDINGS
  Department: Departamento de Representación Arquitectónica (RA)
  Mode: Normal
  Deposit date: 11/05/2026
  Reading date: 16/06/2026
  Reading time: 11:00
  Reading place: ETSAB (Escola Tècnica Superior d'Arquitectura de Barcelona) - Planta Baixa - Sala de GrausAv. Diagonal, 649-651 - 08028 - Barcelona
  Thesis director: NAVARRO DELGADO, ISIDRO | SÁNCHEZ RIERA, ALBERTO
  Thesis abstract: This doctoral thesis examines the layout, evolution, and functions of choirs in Spanish Gothic cathedrals from the 15th and 16th centuries, distinguished by their location in the central nave. Through an interdisciplinary approach combining historical, architectural, and iconographic analysis with advanced photogrammetric techniques, several choir ensembles are digitally documented. The resulting three-dimensional models provide new interpretations of the Spanish choir model, highlighting its liturgical, symbolic, and socio-political significance, while establishing replicable digital methodologies for heritage documentation, preservation, and dissemination.
  

• TORRES LERMA, JOSE ANTONIO: Additive manufacturing constraints in topology optimization using a perimeter functional and a null space algorithm
  
  Author: TORRES LERMA, JOSE ANTONIO
  Programme: DOCTORAL DEGREE IN STRUCTURAL ANALYSIS
  Department: Department of Civil and Environmental Engineering (DECA)
  Mode: Normal
  Deposit date: 19/05/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: FERRER FERRE, ALEX | OTERO GRUER, FERMÍN ENRIQUE
  Thesis abstract: In the context of lightweight structural design, this thesis addresses the incorporation of additive manufacturing constraints into topology optimization in a simple, general, and computationally efficient manner. In particular, the focus is placed on two key limitations arising in additive manufacturing processes: the minimum length scale and overhang constraints. Existing approaches often rely on complex modifications of the governing physics or on additional mechanical constraints, leading to increased computational cost and implementation complexity.To overcome these limitations, this work proposes a unified framework based on regularized perimeter constraints, which can be consistently applied to both density-based and level-set formulations. To the best of the author’s knowledge, this represents the first extension of perimeter-based methods to the local enforcement of additive manufacturing constraints. Nonlinear smoothing extensions are introduced to solve the overhang constraints, while we include the definition of minimum thickness constraints through an isoperimetric analogy. A dual discretization strategy is also developed to enforce the constraints locally.In parallel, an extended null space optimization algorithm is proposed to efficiently handle the resulting multi-constraint problems while requiring minimal parameter tuning. The method is shown to be applicable to density-based approaches, shape optimization, and level-set methods with topological derivatives. Furthermore, two acceleration strategies are investigated - namely, a subiteration approach and a quasi-Newton method - demonstrating improved convergence behavior through the incorporation of nonlinearities in geometrical functionals.The results show that the proposed methodology provides an effective and computationally efficient framework for enforcing additive manufacturing constraints, while maintaining flexibility across different design representations. The combination of perimeter-based constraints and a robust optimization algorithm offers a promising alternative to existing approaches, particularly for large-scale and complex applications.
  

Reading date: 17/06/2026

• FERNANDEZ, LORETTE SYLVIE JACQUELINE: Design and Characterization of More Sustainable Molecular Solar Thermal Energy Storage Systems
  
  Author: FERNANDEZ, LORETTE SYLVIE JACQUELINE
  Programme: DOCTORAL DEGREE IN CHEMICAL PROCESS ENGINEERING
  Department: Department of Chemical Engineering (EQ)
  Mode: Normal
  Deposit date: 20/05/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: MOTH-POULSEN, KASPER
  Thesis abstract: Energy demand and consumption associated with global population growth are driving an energy transition from fossil fuels to more sustainable alternatives, with storage at the core of this revolution. Among the various approaches, molecular solar thermal energy storage (MOST) is gaining increasing attention. The emission-free MOST systems capture solar energy, store it chemically, and release it on demand in the form of heat. This thesis examined three strategies to further promote the sustainability of MOST systems. The first relies on enhancing the performance of a small liquid solar energy-harvesting device by introducing various reflective elements beneath the collector. An increase of 0.1% in solar energy storage efficiency is achieved compared with a non-reflective background. The second strategy focuses on employing surfactants to disperse organic MOST photoswitches in water. Sunlight (simulated) harnessing and macroscopic heat release (under ambient conditions) of 4.7 °C are demonstrated in such formulations. The third examines the optical properties of MOST coatings based on cellulose nanocrystals for the development of solid-state devices. Overall, these strategies pave the way for future sustainable research directions for MOST systems.
  

• MARTÍ ESPELT, ANIOL: Physically consistent wireless communications with statistical channel state information
  
  Author: MARTÍ ESPELT, ANIOL
  Programme: DOCTORAL DEGREE IN SIGNAL THEORY AND COMMUNICATIONS
  Department: Department of Signal Theory and Communications (TSC)
  Mode: Normal
  Deposit date: 20/05/2026
  Reading date: 17/06/2026
  Reading time: 11:00
  Reading place: Aula de Teleensenyament, Edifici B3, Campus Nord, Barcelona
  Thesis director: RIBA SAGARRA, JAUME | LAMARCA OROZCO, M. MERITXELL
  Thesis abstract: The deployment of 5G and 6G wireless networks enforces a paradigm shift in communication system design, driven by the move towards sub-THz frequencies and extremely large antenna arrays. These advancements leave traditional channel modeling—which often relies on far-field assumptions and neglects electromagnetic interactions between antenna elements—physically inconsistent. Furthermore, the overhead associated with acquiring instantaneous channel state information (CSI) in massive multiple-input multiple-output (mMIMO) systems presents a critical bottleneck, particularly for low-latency or high-mobility communications.This thesis addresses the aforementioned challenges by developing a framework for the design and analysis of noncoherent wireless communication systems that operate solely with statistical, rather than instantaneous, CSI. The core of this work is the establishment of a physically consistent channel model that accurately incorporates the effects of near-field spherical wavefronts and mutual coupling. We demonstrate that these complex physical phenomena can be effectively captured within a correlated Rayleigh fading model, providing a tractable yet realistic foundation for system analysis.Using this framework, we investigate the performance of one-shot, energy-based communication schemes, which are particularly well suited for low-latency applications. A key result is the existence of a fundamental error floor at high signal-to-noise ratio (SNR) for constellations with more than two energy levels when no CSI is available at the transmitter. However, we also prove that this error vanishes as the number of receiver antennas grows, highlighting the channel hardening benefits of massive arrays.A widely adopted receiver in energy-based noncoherent systems is the so-called energy detector. Although it is optimal under uncorrelated fading, its performance degrades significantly in correlated channels. To address this limitation, we introduce a novel class of quadratic detectors, including the best quadratic unbiased estimator (BQUE) as well as a practical implementation called assisted BQUE. These detectors leverage statistical CSI to achieve near-optimal performance. Furthermore, two strategies for enhancing reliability are proposed and evaluated: a constellation design methodology that minimizes the analytical symbol error rate by leveraging statistical CSI at the transmitter, and a permutational index modulation (PIM) scheme that introduces coding gain with minimal complexity.Finally, the thesis explores the impact of model mismatch, revealing that noncoherent systems exhibit greater robustness to mutual coupling than their coherent counterparts. We also demonstrate that wavefront curvature can be exploited well beyond the classical Fraunhofer distance. Moreover, we show that large antenna arrays enable the multiplexing and low-complexity detection of multiple users, even when employing noncoherent processing.