ML Training Session for Early Career Researchers

• Timothy Heightman (ICFO)

Room: Aula Magna

ML Training Session for Early Career Researchers

• Timothy Heightman (ICFO)

Room: Aula Magna

Welcome Room: Aula Magna

Interpretable deep learning and quantum machine learning for medicine

• Miguel Ángel González Ballester (Universitat Pompeu Fabra & ICREA)

Room: Aula Magna Deep learning has had a profound effect in all walks of life, not least in medical research. The impressive performance of these methods is partly due to the development of very efficient approaches to solve highly overparameterized neural networks. Conversely, this leads to the problem of trust and interpretability of these complex models, which are key aspects in the medical domain. In this talk, we will see several works performed in our lab focusing on explainability and uncertainty modelling for a variety of clinical applications. We will then delve into the interface between machine learning and quantum computing, and explore early works on the development of quantum machine learning approaches for healthcare. First, we will show some results on the establishment of quantum diffusion models for medical image analysis. Finally, we will link back to explainability, formulating a quantum annealing framework for explainable deep learning

Quantum Scrambling as a resource for quantum generative models

• Marcin Płodzień (Qilimanjaro Quantum Tech)

Room: Aula Magna

Networks and AI to explore materials space

• Guido Caldarelli (CNR-ISC)

Room: Aula Magna

Learning to simulate quantum circuits from samples

• Ryan LaRose (Michigan State University)

Room: Aula Magna

Predicting Quantum Many-Body Dynamics with Classical and Quantum Attention Models

• Matteo Rosati (Università Roma Tre)

Room: Aula Magna Predicting the real-time dynamics of quantum many-body systems is a central challenge in quantum physics, where the exponential growth of Hilbert space makes direct simulation rapidly intractable. In this contribution, we explore attention-based models as a framework for learning and extrapolating quantum dynamics. Attention mechanisms are the core building blocks underlying modern large language models: in language, they identify which previous words are relevant for predicting the next one; here, we use the same idea to identify which previous quantum states are relevant for predicting future time evolution. First, we introduce a variational quantum self-attention mechanism in which future elements of a sequence are generated from overlap-weighted combinations of past data. The required nonlinearity arises naturally from quantum interference between state overlaps, and the training objective can be expressed directly as an observable expectation value. This provides a quantum-native route to sequence prediction and yields a quadratic speedup in sequence length relative to classical attention, while remaining applicable to classical data and transverse-field Ising model trajectories. Second, we present ongoing work showing that classical attention models can learn the structure of quantum time evolution from datasets of trajectories generated within a given Hamiltonian class. Given a short sequence of evolved states up to time t, the model predicts the continuation up to times T≫t. Across TFIM, Fermi-Hubbard, XXZ, Heisenberg, and 3-MAX-SAT Hamiltonians, we obtain high test fidelities and accurate predictions of observables such as magnetization. We discuss key challenges for generative quantum dynamics, including long-time stability, exposure bias, and the representativity of training ensembles. Overall, our results suggest that attention models, both classical and quantum, provide flexible inductive biases for learning and extrapolating complex quantum many-body dynamics. References: https://arxiv.org/abs/2602.06699

Quantum Computing with Words

• Giovanni Acampora (University of Naples Federico II)

Room: Aula Magna

Learning dynamics and Fast Sampling for Energy-Based Models

• Aurélien Decelle (Universidad Politécnica de MadridU)

Room: Aula Magna

Learnability of noise in quantum devices

• Cristina Cirstoiu (Quantinuum)

Room: Aula Magna

Predicting Phonon Thermal Transport in 2D Materials with Machine-Learning Interatomic Potentials

• Cem Sevik (Department of Physics, University of Antwerp, Belgium)

Room: Aula Magna Group-VI transition metal dichalcogenides, such as MoS$_2$ and MoSe$_2$, are prototypical two-dimensional materials with distinctive phononic and electronic properties, making them highly attractive for nanoelectronic, optoelectronic, and thermoelectric applications. However, their reported lattice thermal conductivities remain highly inconsistent, with experimental measurements and theoretical predictions differing by more than an order of magnitude. In this work, we systematically investigate the origins of these discrepancies by combining first-principles calculations, molecular dynamics simulations, and machine-learning interatomic potentials (MLIPs). We employ and benchmark several state-of-the-art models, including GAP, MACE, NEP, and HIPHIVE, against density functional theory to assess their accuracy and transferability [1]. The computational efficiency of MLIPs enables extensive convergence studies that go beyond conventional limits, allowing a rigorous evaluation of higher-order anharmonic effects. In particular, we quantify the role of third- and fourth-order phonon scattering processes and validate our findings using homogeneous nonequilibrium molecular dynamics simulations. Our results demonstrate that, contrary to recent claims, fully converged four-phonon scattering contributes negligibly to the intrinsic thermal conductivity of both MoS2 and MoSe2. This resolves a key source of inconsistency in the literature and establishes reliable benchmarks for these systems. More broadly, this work highlights the potential of machine-learning interatomic potentials as a robust and scalable framework for predictive modeling of phonon-mediated thermal transport in low-dimensional materials. Reference: [1] Tuğbey Kocabaş, Murat Keçeli, Tanju Gürel, Milorad V Milošević, Cem Sevik, Thermal conductivity limits of MoS$_2$ and MoSe$_2$: Revisiting high-order anharmonic lattice dynamics with machine learning potentials, Appl. Phys. Rev. 12, 041424 (2025).

Interpretable machine learning for quantum simulators

• Anna Dawid (Leiden University)

Room: Aula Magna

Neural quantum states through the lens of modern machine learning

• Agnes Valenti (Flatiron Institute)

Room: Aula Magna

Probabilistic learning of closed-form mathematical models from data

• Roger Guimerà (ICREA & Universitat Rovira i Virgili)

Room: Aula Magna

What geometry are graph neural networks learning?

• Marián Boguñá (University of Barcelona & ICREA)

Room: Aula Magna

Neural Quantum Kernels for classification

• Camila Roxana Cristiano Romero (Basque Center for Applied Mathematics)

Room: Aula Magna

Nuclear-physics-guided Gaussian Processes

• Hristijan Kochankovski (Departament de Fisica Quantica i Astrofisica and Institut de Ciencies del Cosmos, Universitat de Barcelona, Marti i Franques 1, 08028, Barcelona, Spain)

Room: Aula Magna Gaussian process (GP) regression is a powerful nonparametric Bayesian method that provides both predictions and principled uncertainty estimates in closed form. While most applications rely on generic priors, the flexibility of the GP framework allows physics knowledge to be systematically encoded into the mean function and the kernel. We exploit this capability in the context of nuclear physics, applying physics-informed GP models to three problems: nuclear binding energies, the finite-temperature equation of state of dense matter, and nucleon-nucleon scattering phase shifts. In each case, we demonstrate that encoding known theoretical structures yields substantial and systematic improvements in interpolation accuracy, uncertainty calibration, and extrapolation reliability over agnostic baselines. Our results highlight that the design of the mean function prior can be as consequential as the choice of kernel, offering a general strategy for physics-informed machine learning of complex quantum systems.

Neural Differential Equations in Quantum Process Tomography

• Timothy Heightman (ICFO)

Room: Aula Magna

Modeling complex systems with energy-based models

• Beatriz Seoane (Universidad Complutense de Madrid)

Room: Aula Magna

Variational learning strongly interacting systems

• Alessandro Lovato (Argonne National Laboratory & IFIC-CSIC)

Room: Aula Magna

Learning the quantum many-body problem: From quantum materials to quantum simulators

• Ivan Morera (ETH Zurich)

Room: Aula Magna

Optimizing Neural Quantum States with Decision Geometry

• Mehdi Drissi (TU Darmstadt, Theory Center)

Room: Aula Magna

Applications of Foundation Neural Quantum States in Many-Body Quantum Systems

• Alessandro Santini (CPhT)

Room: Aula Magna

Efficient estimation of the Partition Function of Restricted Boltzmann Machines via Annealed Importance Sampling

• Ferran Mazzanti (Universitat Politècnica de Catalunya)

Room: Aula Magna

Simulating nuclear physics using gadget Ising hamiltonians

• Emanuele Costa (University of Barcelona)

Room: Aula Magna We present a perturbative gadget construction that maps quasiparticle nuclear shell model (NSM) Hamiltonians onto effective Ising dynamics. Starting from a one-hot encoded quasiparticle basis, we derive an effective low-energy Hamiltonian via Brillouin-Wigner perturbation theory, recovering the target nuclear interactions as emergent couplings of a perturbed Ising model. We discuss how this construction opens a concrete route to simulating nuclear structure Hamiltonians on Rydberg atom platforms and quantum annealers, bridging the gap between many-body nuclear physics and quantum simulators.

A machine learning tool to analyze spectroscopic changes in high-dimensional data

• Giancarlo Franzese (Universitat de Barcelona)

Room: Aula Magna When analyzing spectroscopic data of complex systems, a key challenge is identifying the features that most effectively distinguish their differences, thereby facilitating the study of physical properties and their evolution. A common strategy involves combining multiple measurements to acquire more detailed information; however, this approach does not always enhance analysis quality, especially when additional noise is present. To address this, we developed an unsupervised machine learning (ML) protocol. As a case study, we applied it to multi-component spectral data—including UV Resonance Raman, Circular Dichroism, and UV absorbance—from nanoparticles (NPs) interacting with proteins to form a corona. Understanding the structural evolution of proteins within the corona across different temperatures is crucial for assessing the safety and toxicity of nanotechnology, yet the impact of NP properties on protein conformation remains poorly understood. We focused on fibrinogen, an essential blood plasma protein, at physiological concentrations and examined its interactions with hydrophobic carbon and hydrophilic silicon dioxide NPs. Our method uncovered significant differences in how protein structure responds to temperature in these two scenarios. The unsupervised ML protocol we developed (a) overcomes the challenges posed by the curse of dimensionality, and (b) effectively handles spectral data from diverse sources. It provides a quantitative assessment of protein structural changes upon adsorption, improving understanding of the relationship between protein conformation and NP interactions. This insight could support the development of nanomedical tools for various therapeutic applications. Furthermore, our method is sufficiently general and can be applied to other spectroscopic analyses from different sources. Details are given in [A. Martinez-Serra, G. Marchetti, F. D’Amico, I. Fenoglio, B. Rossi, M. P. Monopoli, and G. Franzese. A machine learning tool to analyze spectroscopic changes in high-dimensional data. International Journal of Biological Macromolecules, 330:148095, 2025].

AI-assisted characterization and tuning of quantum dot-based quantum simulators

• Jarosław Pawłowski (Wroclaw University of Science and Technology)

Room: Aula Magna Quantum simulators based on semiconductor quantum dots (QDs) [1] are promising candidates for practical quantum technologies in the present NISQ (*Noisy Intermediate-Scale Quantum*) era. However, a fundamental challenge is the characterization and control of already fabricated devices. This task becomes increasingly complex in arrays of many interacting QDs, where the number of tunable parameters grows rapidly. The primary experimental tool for device characterization are transport measurements, where systems are tuned by searching for characteristic structures in conductance maps as functions of control parameters such as gate voltages applied to electrostatically defined QDs [2]. Interpreting such high-dimensional data and inferring the underlying Hamiltonian parameters is a difficult *inverse problem*. Machine learning (ML) methods have proven useful for quantum technologies in recent years [3,4]. In this talk we present recent results on ML-assisted analysis and control of QD-based quantum simulators. First, we introduce a physics-informed neural architecture capable of learning effective Hamiltonians directly from conductance maps [5]. Second, we demonstrate an AI-enhanced autotuning protocol that can steer a QD chain toward regimes hosting *Majorana zero modes* by iteratively updating experimentally accessible parameters [6]. Together, these approaches provide a route toward automated characterization and control of complex quantum devices, enabling scalable operation of quantum-dot simulators. **References:** [1] F. Borsoi et al., *Nat. Nanotechnol.* **19**, 21 (2024). [2] J Pawłowski et al., *Nanotechnology* **36**, 195001 (2025). [3] J Pawłowski, M Krawczyk, *Phys. Rev. Applied* **22**, 014068 (2024). [4] M Krawczyk, J Pawłowski, MM Maśka, K Roszak, *Phys. Rev. A* **109**, 022405 (2024). [5] J. Pawłowski, M. Krawczyk, *arXiv*: **2603.02889** (2026). [6] M. Krawczyk, J. Pawłowski, *arXiv*: **2601.02149** (2026).

Dissipation–Performance Interplay in Quantum Reservoir Computing

• Christopher Gies (Carl von Ossietzky Universität Oldenburg)

Room: Aula Magna Quantum Reservoir Computing (QRC) has emerged as a powerful paradigm for dynamical quantum machine learning that is particularly suited for the noisy intermediate\-scale quantum (NISQ) era where gate fidelities remain a limiting factor. As recent experimental and theoretical advances have demonstrated that QRC can leverage the intrinsic dynamics of quantum systems to process temporal information, we need a comprehensive understanding of how physical properties constrain or enhance computational performance. We address this question by establishing a direct link between the optical absorption spectrum of a QRC platform, and its linear and nonlinear information processing capacities. The absorption maximum identifies the operational regime in which the reservoir makes the input signal most effectively available within its dynamical degrees of freedom. This connection provides a physical interpretation of established and commonly used computational QRC benchmarks and enables experimentally tunable reservoir design by engineering dissipation rates and absorption characteristics [1]. Even more intriguing, this behavior is not platform specific, but emerges from a more general underlying tradeoff between memory retention and non-linear performance measures across diverse QRC implementations. Our framework unifies prior observations based on phase transitions, weak measurements, and dissipation‑driven dynamics, demonstrating that these mechanisms enhance performance through the same fundamental principle [2]. [1] N. Götting, S. Wilksen, A. Steinhoff, F. Lohof, and C. Gies, *“Connection between Memory Performance and Optical Absorption in Quantum Reservoir Computing,”* Phys. Rev. Lett. **135**, 240403 (2025). [2] S. Cindrak, L. Giebeler, N. Götting, C. Gies, and K. Lüdge, *“Memory-Nonlinearity Trade-off across Quantum Reservoir Computing Frameworks,”* arXiv:2603.21371 (2026).

Entanglement-assisted Hamiltonian dynamics learning

• Ayaka Usui (Universitat Autònoma de Barcelona)

Room: Aula Magna Approximating the dynamics given by a complex many-body Hamiltonian with a simpler effective model lies at the interface of quantum Hamiltonian learning and quantum simulation. In this context, quantum generative adversarial networks (QGANs) have been shown to outperform standard Trotter-based approximations. However, their performance is often hindered by training plateaus and local minima that become increasingly severe with system size. To overcome these limitations, we propose an entanglement-assisted learning strategy that couples a single randomly initialized auxiliary qubit to the learning system at an intermediate stage of the training process [1]. The interplay between randomization and entanglement significantly enhances the learning performance of the protocol. [1] arXiv:2602.15931