We cordially invite you to the 3rd International Workshop on Machine Learning and Quantum Computing Applications in Medicine and Physics, which will take place in Warsaw (Poland) from 7 to 11 September 2026. The workshop is organized by the National Centre for Nuclear Research in cooperation with scientists from the University of Vienna, Jagiellonian University.
In the last years, we have seen an increasing number of interdisciplinary exchanges, especially with machine learning methods involved. Many times similar techniques and tools are used to solve a variety of problems. The transfer of methods between disciplines is highly needed and it has repeatedly proved to be a very fruitful approach. This workshop aims to facilitate such transfers of knowledge by bringing together experts from various institutions and research groups in the fields of medical imaging, radiotherapy, physics, and computer science. The workshop serves as a platform to enable research at the edge of various disciplines.
The scope of the workshop covers among others quantum simulations, quantum algorithms, and (classical or quantum) machine-learning algorithms with a focus on application in Physics and Medicine.
List of topics:
Feel free to share this invitation with your peers and whoever you think would be interested. Please, do not hesitate to contact us in case of any questions: wmlq2026@ncbj.gov.pl
Links to previous WMLQ editions:
Short communication from the workshop organizers
With the advent of the Vera Rubin Observatory, astronomy and astrophysics are entering an era of indeed astronomically big data. I will discuss new opportunities but also challenges that machine learning is facing today, illustrated by examples of research conducted by our Warsaw group.
The era of wide-field spectroscopic surveys such as WEAVE, DESI, and 4MOST, along with forthcoming facilities like the Wide-Field Spectroscopic Telescope (WST) and MOSAIC, is delivering millions of spectra that encode the chemical and structural history of the Universe. The sheer volume of this data makes traditional human-supervised analysis intractable and necessitates the transition to fully automated machine learning (ML) pipelines to efficiently characterize the Intergalactic Medium (IGM), the circumgalactic medium (CGM), and complex galaxy and AGN physics.
A fundamental challenge in spectroscopic analysis is the accurate estimation of the intrinsic spectral continuum. We present an optimized autoencoder architecture that achieves superior precision, with a median Absolute Fractional Flux Error (AFFE) of 0.009 for quasar spectra. This model has demonstrated strong generalizability by successfully recovering the Lyα optical depth evolution in unseen DESI data.
Furthermore, specialized U-Net architectures are now capable of the generalized detection of metal absorption features (e.g., CIV, MgII, SiIV) with high completeness and purity, reaching an F1 score of ≈90% at S/N≈4. Following detection, physically motivated classifiers allow for robust ion identification with an average 90% accuracy, reaching ∼100% for reliable doublet systems. This integrated pipeline allows the analysis of ≈10,000 spectra in just a few seconds.
Galaxy mergers are fundamental to the hierarchical assembly and evolution of galaxies, often driving starburst activity and AGN fueling. Identifying mergers and their stages, such as pre- and post-coalescence, from imaging alone, especially given the vast size of modern datasets, remains extremely challenging. We develop a supervised deep learning framework using Convolutional Neural Networks (CNNs) to classify galaxies as non-mergers, pre-mergers, or post-mergers. Our training uses mock Hyper Suprime-Cam (HSC) images from the IllustrisTNG simulations (Margalef-Bentabol et al. 2024). HSC, a precursor to LSST, is ideal for developing and validating machine learning methods for future surveys. We test our model on synthetic and real HSC data to assess robustness and generalizability. This approach demonstrates the potential of simulation-driven machine learning to reveal galaxy merger histories in upcoming wide-field surveys.
The rapidly increasing demands of LHC physics pose a major challenge for Monte Carlo simulation. With the High-Luminosity LHC expected to deliver an order of magnitude more data, generating sufficiently large and high-fidelity simulated datasets will become increasingly computationally demanding. Recent developments demonstrate that generative machine learning can provide substantial computational speed-ups while retaining the accuracy required for physics analyses.
However, the potential of generative simulation goes beyond simply replacing an expensive simulation step. Once a generative model is trained on a finite Monte Carlo dataset, it can be sampled arbitrarily many times, raising a fundamental question: how many of these generated events are statistically useful? The concept of generative amplification addresses precisely this question by comparing the statistical power of generated samples with that of the original training data.
I present a framework for quantifying generative amplification without requiring access to large independent holdout samples. Two complementary approaches are introduced: an averaging measure based on phase-space integrals and a differential measure based on hypothesis testing. Applied to generative models for LHC event generation, these methods allow us to identify regions of phase space in which generated samples can provide genuine statistical amplification, as well as regions where model uncertainties limit the effective sample size.
Fast simulation of calorimeter showers is critical for the High-Luminosity LHC. Traditional Monte Carlo methods are computationally prohibitive for Run 4 of the LHC. We present DSSGAN3D (Directional State Space GAN), a generative adversarial network for volumetric simulations evaluated on the CaloChallenge benchmark (electron showers).
The GAN generator is conditioned on layer-wise energy fractions from a low-dimensional conditional flow matching (CFM), following the established two-stage paradigm. Our parallel CFM Transformer integrates one joint 45-dimensional ordinary differential equation, generating all fractions simultaneously.
The main GAN shower generator employs a bidirectional multi-axis Mamba backbone over the radial-longitudinal subvolume. The azimuthal coordinate lacks sequential structure and is treated as independent channels, with no cross-phi interaction in the SSM. A final lightweight PhiAttention head jointly generates all phi outputs at each position via self-attention. Conditioning via Directional Latent Routing decomposes the latent into axis-specific subvectors modulated by the energy embedding per scan direction. The generator uses full-resolution 2D spatial noise, matching the output grid directly.
DSSGAN3D achieves FPD=0.0557 with our parallel CFM model, the second-lowest among original CaloChallenge entries, surpassing all except CaloDREAM. The full pipeline runs nearly two orders of magnitude faster than other top-ranked submissions.
Designing the next generation colliders and detectors involves solving optimization problems in high-dimensional spaces where the optimal solutions may nest in regions that human experts would normally not explore. Meanwhile, the staggering simulation demands of existing and future high-energy physics facilities call for a new paradigm for event generation and reconstruction.
Differentiable programming offers a path forward. By integrating domain knowledge encoded in simulation software with gradient-based optimization and reinforcement learning, it enables end-to-end experimental design and inference in settings that are intractable with conventional methods.
In this talk I will describe recent results for the AI-assisted optimization of experimental design, with a focus on large-scale simulation software, touching on recent advances in calorimetry with neuromorphic hardware architectures, and on medical applications, paving the way to more complex challenges.
High-fidelity Monte Carlo simulations (e.g., Geant4, CORSIKA, Pythia) are indispensable in high-energy and medical physics. However, aligning simulated and experimental distributions remains a laborious task, traditionally relying on manual selection tuning or computationally expensive grid scans. This challenge is compounded when quality selection cuts must be jointly optimized across a high-dimensional phase space containing correlated detector observables.
An automated, domain-independent framework designed to efficiently optimize multi-variable quality selection criteria is introduced. Utilizing a pure-Julia architecture leveraging automated optimization engines, the pipeline simultaneously minimizes binned shape discrepancies while applying density-based regularizations to preserve overall statistical efficiency. The methodology is evaluated using both synthetic and publicly available data. The high performance of Julia's compilation ecosystem enables end-to-end multi-dimensional optimization over millions of events in seconds. Foreseen future applications of this framework include the optimization of complex data pipelines for large-scale neutrino observatories like KM3NeT, as well as next-generation medical workflows such as the 3Dπ liquid argon PET scanner. The resulting open-source toolkit provides an effective approach to simulation tuning for both complex physical event analyses and precision medical imaging workflows.
Electron-impact processes are fundamental to analytical chemistry, plasma physics, radiation science, and astrochemistry, yet experimental measurements remain costly and incomplete. We present machine learning approaches for predicting two key electron-impact observables directly from molecular structure: electron ionization mass spectra (EI-MS) and total electron-impact ionization cross-sections.
For EI-MS prediction, we developed a hybrid model combining a graph neural network encoder, a residual neural network decoder, cross-attention refinement, bidirectional prediction, and chemistry-informed masking. Trained on the NIST14 EI-MS database (molecules ≤500 Da), the model achieves strong library-matching performance (Recall@10 ≈ 80.8%). For ionization cross-sections, we introduce a stacked ensemble integrating LightGBM, CatBoost, a multilayer perceptron, and a graph neural network, combined through a residual LightGBM meta-model. Trained on 212 molecules from the Astrochemistry Low-energy Electron Cross-Section database, the ensemble achieves a test-set RMSE of 1.160 a₀² and a Spearman correlation of 0.990, outperforming all individual models.
The results demonstrate that graph-based molecular representations enable accurate prediction of complementary electron-impact properties. These approaches can augment reference databases, support molecular identification, and reduce the need for costly experimental measurements.
AI applications are currently hot topic in the field of HPC and require significant computing resources. LUMI supercomputer, currently positioned 11th on the Top500 list, hosts over 3000 scientific projects, and over half of LUMI’s computing resources have been utilized for AI-related research and innovation making it one of the world’s most powerful AI platforms for science, playing important role in supporting European research community.
Looking into the future of HPC, the EuroHPC Joint Undertaking has selected the hosting sites of the next EuroHPC supercomputers and AI Factories. One of the chosen hosting sites is Finland, led by CSC – IT Center for Science, together with a LUMI AI Factory consortium of five other countries: the Czech Republic, Denmark, Estonia, Norway and Poland.
An AI Factory is an ecosystem that enables AI researchers and developers to have one-stop access to the high-performance computing, data sets and skills they need. The aim is to make it as easy as possible for both scientific researchers and industrial innovators to adopt AI methods on a large scale.
In this talk, I will present the architecture of the LUMI infrastructure and its status, together with plans and ambitions for the near future. Then I’ll give examples of computations done on modern HPC, both with and without AI. Finally, I’ll discuss the applications of high-performance computing in physics, in particular, prospects of benefiting from the Nordic AI Factories.
The National Institute for Nuclear Physics (INFN) is a distributed research institute whose computing resources span its federation. Through the AI_INFN initiative and national projects such as TeRABIT and ICSC, INFN has built a large-scale infrastructure for Machine Learning, combining GPU clusters, HPC systems, FPGA nodes and high-performance storage. It serves a growing community applying ML to physics research, with a focus on scalability, open science and emerging paradigms such as quantum computing.
Building on the architecture developed at the CNAF Tier-1 site, which optimized hardware orchestration and the user experience for interactive and batch workloads, the ReCaS-Bari site within the INFN Cloud federation has been selected for platform replication and functional extension. This deployment proves the design replicable across data centres and serves as a testbed for frontier features, including workload offloading to the Leonardo supercomputer and other HPC centres, already validated from CNAF with Argo Workflows dispatching pods through interLink. Establishing ReCaS-Bari as an alternative to the CNAF instance lays the foundation for cross-site federation and a geo-distributed AI computing infrastructure within INFN. Multi-tenant storage and POSIX object-storage access are in production, while offloading from ReCaS-Bari and Kubernetes accounting are under development, on a backbone open to computing centres, research groups and scientists across disciplines.
We describe main application areas served at Świerk Computing Centre, a supercomputing facility of NCBJ. These include nuclear reactor simulations and design, high energy physics, astrophysics, material science, FEL science, chemistry and drug design, power grid simulations - HPC supplemented with AI for most of the mentioned cases. These applications implement completely different computing models and thus have different, often contradicting architectural requirements. Here we show how we deal with these conflicts, while trying to maintain elasticity in resource usage and reasonable investment and operational costs.
I will discuss how concepts and tools from quantum information theory, in particular the notion of entanglement, can be applied to elementary particle scattering at high-energy colliders. In this framework, the discrete quantum numbers of scattered particles are identified with quantum-information-theoretic qubits, and the resulting density matrix reflects the structure of the underlying perturbative expansion. I will review some recent findings that connect the entanglement properties of scattered particles to specific features of scattering amplitudes. Finally, I will discuss how minimization of quantum entanglement generated in a scattering event can lead to the emergence of symmetries.
Recently, quantum machine learning, particularly quantum reservoir computing, has attracted attention for directly utilizing the natural complex dynamics of quantum systems for information processing. It is believed that by utilizing the vast internal space (Hilbert space) of quantum systems, performance comparable to classical models can be achieved with only a few qubits. However, encoding classical information into a small number of qubits makes it challenging to apply quantum models to large inputs, such as image classification. In this context, in 2022, we proposed a new quantum model, quantum extreme reservoir computing (QERC), that combines classical lightweight compression techniques and achieves performance comparable to classical models within the same machine learning family. This talk presents the QERC model and the various quantum reservoir systems we have experimented with. These include a wide range of possibilities, from using natural Hamiltonian dynamics, as in the Ising model, to representing quantum circuits with random Clifford circuits. Furthermore, this talk will address how much one can simplify the reservoir and how much complexity is necessary, which is of interest to both the theory and implementation sides.
AI now rivals or surpasses humans on many well-defined tasks, sharpening the need to clarify what "information" means across disciplines. This talk frames information operationally: Shannon entropy for classical systems and von Neumann entropy for quantum states, and shows how these measures structure, analyze, and reduce uncertainty in machine learning and data-driven inference. We compare classical and quantum information—highlighting superposition, entanglement, measurement disturbance, and no-cloning—and connect these principles to technologies ranging from Positron Emission Tomography (PET), where information limits govern reconstruction quality and dose–time trade-offs, to quantum computing, where information is encoded, processed, and
protected under fundamentally different constraints. We conclude with a speculative platform that illuminates the interface of gravity and quantum control: using the discrete gravitational bound states of a single neutron as a qudit. The discussion emphasizes the unifying role of information across disciplines.
Medical imaging provides a demanding test for quantum machine learning because clinically relevant images are much larger and structurally richer than standard benchmarks (MNIST). In this work, we study quantum extreme reservoir computing (QERC) for binary brain-tumour classification using the Br35H MRI dataset, with images resized to 150 × 150 pixels. Since such inputs cannot be encoded directly into a small quantum reservoir, we compare principal component analysis with two autoencoder-based compression schemes.
The results show that reconstruction quality alone is not a reliable guide to downstream quantum classification. An autoencoder can reproduce the images well while mapping them into a narrow region of latent space, thereby supplying the quantum encoder with insufficiently varied inputs. We quantify this concentration using the participation ratio and find that broader latent distributions generally produce better and more stable classification. Batch normalization improves performance by spreading the latent variables more evenly before quantum encoding.
These results show that the compression method is not simply a preprocessing choice. For this type of quantum image-classification pipeline, it must both retain useful information and generate latent variables that can be encoded effectively by the quantum reservoir.
Generative models provide powerful learned priors for solving challenging inverse problems in medical imaging. In this talk, I will present diffusion-based, physics-informed approaches for CT and PET reconstruction, illustrated through three applications: motion-compensated head cone-beam CT, material decomposition in photon-counting CT, and joint activity-attenuation reconstruction in CT-less PET. These examples demonstrate how learned image priors can be combined with measurement physics to achieve high-quality, data-consistent reconstruction without paired training data.
We present a compact 3D neural network for segmentation of maxillary sinuses from CT scans in children with chronic rhinosinusitis. It comprises 12 convolutional layers totaling 50k parameters, minimal compared to most 3D medical-imaging architectures, and completes the prediction in a few seconds on an average CPU, making the tool accessible for most clinicians. The sinus volume and its fraction occupied by inflammatory changes are estimated with an RMSE of 0.5 cm$^3$ and 1.5 %pts, respectively, which allows for precise longitudinal monitoring. The network includes no dropout, no pooling, and no padding, the latter preventing layer-wise injection of meaningless zeros. We utilize a custom normalization that does not average the data but employs the same running statistics in both prediction and training, making the model insensitive to overall contrast and local artifacts. The network is fully convolutional and translation-invariant. It has an inner receptive field of 18x18x18 voxels to detect sinus walls and an outer receptive field of 88x88x88 voxels to provide a broader context. The internal data flow is designed to minimize the number of mappings that the network must learn. It operates in a reduced resolution of 1x1x1 mm$^3$, which still allows for high precision thanks to the use of fuzzy labels accounting for the partial-volume effect. The model was trained and tested on a dataset of 92 scans collected and manually annotated specifically for this study.
Data-driven reconstruction for computed tomography has boomed in the last decade. Yet unlike in other AI fields, its adoption in practical CT has been slow. In this talk I'll give a brief overview of the types of reconstruction methods and highlight the current challenges and some progress that we have done recently
We introduce a novel learning framework for accelerated Monte Carlo (MC) dose calculation termed Energy Shifting. This approach leverages deep learning to synthesize complex polyenergetic dose distributions directly from simple monoenergetic inputs under identical beam configurations. Unlike conventional approaches such as denoising or GenAI techniques, our method achieves superior cross-domain generalization on unseen datasets by integrating high-fidelity anatomical textures and source-specific beam similarity directly into the model’s input space. Furthermore, we propose a novel 3D architecture termed TransUNetSE3D, featuring Transformer blocks for global context and Residual Squeeze-and-Excitation modules for adaptive channel-wise feature recalibration. Hierarchical representations of these blocks are fused into the network’s latent space alongside the primary dose-map parameters, allowing physics-aware reconstruction. This hybrid design outperforms existing U-Net and Transformer-based benchmarks in both spatial precision and structural preservation, while maintaining the execution speed necessary for real-time use. Our proposed pipeline achieves a Gamma Passing Rate exceeding 98% (3%/3mm) compared to the MC reference, evaluated within the framework of a treatment planning system using 6MV TrueBeam Linear Accelerator for prostate radiotherapy. These results offer a robust solution for fast volumetric dosimetry in adaptive radiotherapy.
Long axial field-of-view PET scanners provide a significant gain in sensitivity, but also bring a substantially higher background of scattered and accidental coincidences. We investigate machine-learning classification of individual coincidence events using GATE Monte Carlo simulations of the Siemens Quadra scanner with NEMA IEC and anthropomorphic XCAT phantoms, where every event has a ground-truth label.
We evaluate XGBoost, AdaBoost, and neural network classifiers with two feature sets, and show that phantom-wise metrics give a misleading picture of classifier performance. In a cross-phantom test, a small decrease in accuracy corresponds to regionally concentrated degradation that only per-voxel quality maps reveal, while geometry-dependent features overfit the phantom geometry.
In our initial approach, we kept events predicted as true and rejected all others. We present two extensions in which the per-event probabilities influence the reconstructed image and its uncertainty. First, using soft probability weighting, in which each event contributes to reconstruction according to its estimated true-coincidence probability, we reduced the image-wise reconstruction error compared to hard filtering. Second, we apply split conformal prediction, which replaces a single predicted class with a set of classes carrying a finite-sample coverage guarantee. Aggregated per voxel, these sets produce ambiguity maps that identify unreliable image regions without ground-truth labels.
Background and Aims: This study introduces graph neural networks (GNNs) to improve prompt-gamma imaging for a 1D coded-mask gamma camera developed by the SiFi-CC collaboration for proton therapy verification.
Methods: The coded-mask detector features LYSO:Ce,Ca scintillating fibers with dual-ended readouts. Coincident detector signals are aggregated into event representations and subsequently into cluster representations to capture spatial, temporal, and energy information. A graph-based learning model infers interaction positions and energies from these representations, feeding the output into a Maximum-Likelihood Expectation-Maximization (MLEM) algorithm for depth profile reconstruction. Finally, image reconstruction parameters are optimized using an NSGA-III multi-objective genetic algorithm within the Optuna framework.
Results: For a dataset of 10$^{10}$ protons, the optimized pipeline demonstrated sub-millimeter beam range accuracy, achieving a root-mean-square error (RMSE) below 0.6 mm.
Conclusions: Combining graph-based event reconstruction with targeted image optimization significantly improves the resolution of coded-mask prompt-gamma imaging detectors.
Medical and industrial imaging have long influenced one another, from CT's migration into nondestructive testing to shared detector technologies. This talk explores that exchange through IMPET, a project developing AI-enhanced, quantum-informed multiphoton imaging for industrial applications such as opaque flow characterisation and 3D porous-media analysis. We present recent progress on positronium-based image reconstruction and simulation, alongside a brief look at our PEPT extension.
Industrial computed tomography allows for non-destructive inspection of complex components,
making it an interesting domain of applications for machine learning methods.
In this context, a large bottleneck for training robust models is data acquisition.
Already time-intensive task of obtaining data using real-world CT scanners
is further exacerbated by the requirement of creating accurate labels.
We solve this problem by implementing a pipeline for the generation of massive
synthetic datasets, and using only some examples of real-world data.
Our data pipeline uses parametrized CAD geometry to generate thousands of
randomized objects from which we derive both ground-truth labels and
attenuation data further fed into tomographic simulations.
The simulation code is GPU-accelerated allowing us to substantially
reduce the time needed to generate the datasets.
To reproduce effects visible in real data, we implement a polychromatic beam,
and simulate artefacts such as non-uniformity of the beam and detector matrix response.
We also apply custom beam hardening corrections for our reconstructions.
We use the synthetic data in semi-supervised training of a W-net model for the specific
task of segmentation of ionization chambers. We show, that semi-supervised approach
allows for effective training with only partial labeling (as low as 25%-50% of the data).
We then validate performance of segmentation models trained using simulations on the real-world data.
The performance of modern, high-load information systems significantly depends on the efficiency of cache management. Meanwhile, the traditional Least Recently Used and Least Frequently Used policies are based on fixed heuristics, which limits their ability to adapt to changes in query structure and intensity.
To overcome these constraints, a decision-making approach for adding new objects to the cache is proposed, which is based on the tabular Q-Learning method with a deferred asynchronous update mechanism. It takes into account the key characteristics of objects, the context of their use and long-term consequences. The proposed approach involves training the agent through the accumulation of experience and the consideration of deferred rewards, which ensures an adaptive caching policy in real time and improves its efficiency, in particular by increasing the cache hit ratio.
The practical value of the results lies in the ability to improve the efficiency of Clinical Decision Support System users when working with current data. These systems process large, rapidly updated sets of medical data, and the developed approach will reduce data access latency and improve the information support provided to doctors when forming clinical recommendations. The proposed approach can also be applied to other electronic health systems for making decisions regarding the caching of frequently used data.
Proton CT is a promising alternative to X-ray CT for proton therapy treatment planning, allowing direct estimation of the relative stopping power map within the patient body without relying on conversion from Hounsfield units. Conventional list-mode proton CT scanners measure the energy loss of each individual proton to estimate the integral of the relative stopping power along its path, the water-equivalent path length, but do not meet the requirements for clinical use due to their low acquisition rates. An alternative proton CT scanner design was recently proposed where the time-of-flight of each proton is measured between two detectors located before and after the patient along the proton beam. The main advantages of this sandwich time-of-flight design are its compactness and high acquisition rates. However, conversion from sandwich time-of-flight data to water-equivalent path length is not possible. We present an iterative algorithm that leverages PyTorch's automatic differentiation engine to directly optimize the voxels in the image space. The method is assessed and compared using Monte Carlo simulations.
Quantum machine learning asks whether quantum computers can offer machine learning capabilities beyond classical methods. Assessing this potential requires a firm understanding of its fundamental concepts, the basic models, and their limitations. This talk reviews these foundations and the challenges of working with quantum devices before focusing on supervised quantum machine learning with variational quantum circuits (VQCs). We discuss how to encode classical data into quantum models and how to train their parameters using gradient-based optimization. Two VQC-based architectures serve as case studies: data re-uploading and dissipative quantum neural networks. We conclude with some central questions facing the field, including the trainability and scalability of variational models.
Nonlocal games provide a framework for studying cooperative strategies under different resource assumptions. These range from independent agents with no shared resource, through local hidden variable (LHV), to entangled quantum states, allowing one to quantify the advantage of quantum strategies over classical ones. Discovering optimal quantum strategies, however, remains a non-trivial task. We propose a multi-agent reinforcement learning (MARL) approach with shared variational quantum circuits (VQCs) as a general-purpose tool for automated discovery of such strategies. Prior to the game, the agents may agree on a strategy, but once the game begins, each agent receives a question and must provide an answer without communicating. An entangled state is prepared across the agents' respective registers. Each agent then applies local parameterized unitary operations conditioned on their received question, followed by a measurement on their subsystem. The measurement outcomes determine the agents' actions, and the circuit is trained via the REINFORCE algorithm. We evaluate the method on three nonlocal games: the CHSH game, the Magic Square game, and the Rendezvous problem. Our results demonstrate that the proposed approach consistently exceeds LHV bounds and achieves winning probabilities matching quantum optimal limits.
Detecting statistical dependence between stochastic processes is a core primitive of causal discovery for dynamical systems. State-of-the-art tests compare whole trajectories with the signature kernel, which propagates a static point-similarity kernel through a Goursat PDE and feeds the resulting Gram matrices into kernel independence tests (HSIC, SDCIT). We investigate replacing the classical RBF static kernel with a quantum embedding kernel: a data-reuploading circuit maps each time-augmented path point to a quantum state, and similarity is the state fidelity, read out with the adjoint-circuit trick. Since only the static kernel is swapped, performance differences are attributable to the kernel itself. The circuit angles and an input scale are trained to maximize a standardized-HSIC proxy of test power using SPSA with Adam and held-out validation, avoiding differentiation through the PDE solver. In simulations on linear SDEs, the trained quantum kernel improves HSIC test power over both the classical and the untrained quantum kernel at small sample sizes, while preserving type-I error control, and supports PC-style recovery of multivariate causal graphs with conditional tests. We further study register widths of 4-12 qubits, connecting fidelity concentration to trainability, and note that the ansatz maps natively onto IBM Heron hardware (CRZ as native RZZ plus virtual RZ; 12-qubit heavy-hex loops).
This training provides an accessible introduction to quantum computing, covering both photonic and gate-based (universal) systems. Through hands-on sessions and live demonstrations, attendees will explore the current state and future potential of quantum technology. By the end of the training, attendees will be able to:
Target audience: Anyone curious about quantum computing, regardless of prior experience.
Requirements: Basic computer knowledge and familiarity with any programming language.
Required tools: Bring your own laptop — a remote Jupyter Notebook environment with everything needed will be provided.
This training provides an accessible introduction to quantum computing, covering both photonic and gate-based (universal) systems. Through hands-on sessions and live demonstrations, attendees will explore the current state and future potential of quantum technology. By the end of the training, attendees will be able to:
Target audience: Anyone curious about quantum computing, regardless of prior experience.
Requirements: Basic computer knowledge and familiarity with any programming language.
Required tools: Bring your own laptop — a remote Jupyter Notebook environment with everything needed will be provided.
Gaussian Boson Sampling (GBS) is a photonic sampling process whose output probabilities can be related to graph-theoretic problems, including clique finding. In this talk, I will discuss the application of GBS to the band-selection problem in hyperspectral satellite data, where the goal is to identify a reduced set of spectral bands while preserving the relevant information.
Although GBS was originally formulated as a photonic process, it is also possible to explore its implementation on qubit-based quantum architectures, which are currently more widely accessible. I will present our recent construction that enables such a mapping. Finally, I will discuss the advantages and limitations of employing GBS to satellite data analysis.
A more general version of the “plane section” algorithm from [Computer Physics Communications, 319:109913, Feb. 2026] is used to represent probability distributions for likelihood optimization. The algorithm from [Computer Physics Communications, 319:109913, Feb. 2026] was implemented in python using the pytorch library allowing gradients to be calculated with respect to the parameters of the representation. This in turn made it possible to perform maximum likelihood estimations using gradient methods, for example the Adam optimizer. Applications of the previous implementation were limited to two dimensions. The new version, described in the talk, can in principle be used with an arbitrarily large number of dimensions. This number is of course subjected to a limit determined by the available computing resources. The new algorithm is also much simpler conceptually opening the possibility of creating implementations in other programming languages. Additionally, the proposed algorithm might allow the analytical calculation of gradients for use in custom automatic differentiation functions. The implementation is tested using two and three dimensional distributions related to the problem of finding charged particle tracks in gas detectors of high energy physics experiments.
Monte Carlo particle transport is a cornerstone of simulation in medical imaging and radiotherapy, but its computational cost can limit its use. Artificial intelligence offers several routes to accelerate these simulations, from denoising low-statistics results and predicting dose distributions to learning detector responses and generating particle phase spaces.
This talk surveys these approaches through medical applications, including nuclear imaging, radiotherapy dose calculation, and optical photon transport in radiation detectors. It examines what each model learns, which parts of the simulation it replaces, and how its output can be validated. Particular attention is given to the distinction between reproducing an average response and preserving the probability distributions and correlations needed for reliable simulation.
The discussion also extends to AI-assisted radiotherapy planning, where a predicted dose distribution must be converted into a physically deliverable treatment plan. Together, these examples highlight the opportunities for combining AI with Monte Carlo methods and the importance of evaluating the complete workflow, beyond prediction accuracy or computational speed alone.
Positron Emission Tomography (PET) systems can measure the decay properties of ortho-positronium (oPs), an intermediate bound state often formed during positron-electron annihilation that decays to photons after a short lifetime. The lifetime and decay kinematics of oPs can be used to probe material properties, have potential to serve as a novel biomarker in disease, and can be used test the Standard Model. Additional sources of charge conjugation-parity (CP) symmetry violation are required to explain the observed matter-antimatter imbalance in the Universe, and CP violation can be constrained by measuring decays under the reversal of an applied magnetic field. A dedicated physics measurement platform was constructed from the architecture of the NeuroSphere brain PET insert for 7-T MRI, including a PET detector array, positronium target with lifetime trigger, and motorized gantry for control of the system orientation inside the 7-T environment. GATE simulations were used to inform the development of a custom data analysis pipeline for oPs event selection and multi-coincidence processing. Systematics were mitigated by combining runs with various target and detector positions (to average out artificial asymmetries caused by assembly or detector efficiency) and by applying a baseline correction from events in a kinematic region with vanishing analyzing power (to account for drift and complex field- and material-related effects). decays in polyvinyltoluene (PVT) were measured at 7 T, finding a long lifetime component near 100 ns and CP violation consistent with zero. This work demonstrates the unique opportunity provided by PET/MR instrumentation to perform high-field physics measurements and proof-of-concept for measuring positronium decays with the full NeuroSphere system.
We explore the adaptive algorithm in Positronium Lifetime Imaging (PLI), a new technique that analyses the local behaviour of positronium (Ps) – a quasi-stable electron-positron ($e^−e^+$) compound [1]. Voxel-dependent Ps lifetime spectra can be acquired using Positron Emission Tomography (PET) with specific $\beta^+\gamma$ sources that emit an additional prompt gamma photon. But mostly due to a need for 3-photon coincidence data, the sensitivity is much lower than in standard PET [2].
Based on the previous development of PLI algorithms based on non-linear fitting of noisy voxel-dependent Ps lifetime spectra [3], we further address the challenge of low-count data. A multi-variable Bayesian penalisation factor is added to the negative log‑likelihood minimisation, using prior information from the approximate low-resolution image. There is a major issue that the multi-channel Ps decay model tends to overfit and converge to local minima. A proposed solution is a multi-start optimisation with random initial guesses, split into two stages – each with different variables fixed or penalised from priors.
We also conduct an inferential analysis of the initial guess distributions and their adjustment to the regions that converge best. That allows for a decrease in their minimal number for successful minimisation and for a performance boost of the PLI algorithm.
[1] Deutsch M. Phys.Rev. 82 455 (1951)
[2] Huang B et al. Comm.Phys. 8 1 (2025)
[3] Shopa RY, Dulski K. BAMS 19 54 (2023)
This training provides a practical introduction to optimizing PyTorch-based AI models on GPUs, with a strong focus on performance profiling and custom Triton kernels. The aim is to equip attendees with the skills needed to identify performance bottlenecks and accelerate GPU-based computations. By the end of the training, attendees will be able to:
Target audience: Users who already work with PyTorch and want to accelerate and optimize their GPU-based numerical or AI computations.
Requirements: Working knowledge of PyTorch tensors and basic GPU concepts; Python proficiency; familiarity with undergraduate-level linear algebra.
Required tools: Bring your own laptop — a remote Jupyter Notebook environment with everything needed will be provided.
This training provides a practical introduction to optimizing PyTorch-based AI models on GPUs, with a strong focus on performance profiling and custom Triton kernels. The aim is to equip attendees with the skills needed to identify performance bottlenecks and accelerate GPU-based computations. By the end of the training, attendees will be able to:
Target audience: Users who already work with PyTorch and want to accelerate and optimize their GPU-based numerical or AI computations.
Requirements: Working knowledge of PyTorch tensors and basic GPU concepts; Python proficiency; familiarity with undergraduate-level linear algebra.
Required tools: Bring your own laptop — a remote Jupyter Notebook environment with everything needed will be provided.