2nd SMASHING Workshop

19 Oct 2026, 09:00 → 23 Oct 2026, 14:15 Europe/Ljubljana

Lanthieri Mansion, Vipava

This workshop is the second network meeting of the SMASH project (here you can find the first edition https://indico.ung.si/event/35). SMASH is a multidisciplinary program centered on developing cutting-edge Machine Learning (ML) and Artificial Intelligence (AI) applications for science and humanities. These include climate science, precision medicine, fundamental physics and linguistics.  It is co-funded by the European Union via the Marie Skłodowska-Curie COFUND action and by the Republic of Slovenia and European Union from the European Regional Development Fund.  SMASH connects scholars from five top-level institutions in Slovenia with 52 associated partners, Slovenian businesses and academic institutions globally.

The Second SMASHING workshop will gather scientists working in the SMASH research areas with the aim to create a multi-disciplinary environment that will foster knowledge exchange between different fields and between academia and industry, thereby building the SMASH community. The workshop will be structured around discussions of different classes of ML/AI techniques, followed up with examples on the successful application in SMASH research areas. More specifically the workshop will focus on:

• Agentic AI approaches to science
• Foundation models
• Generative models and computer vision
• Statistical approaches (simulation based inference, etc) 
• Graphs/transformers and time series

In addition to invited talks the workshop will have discussion sessions and there will be plenty of time to exchange ideas and build the community. 

Speakers include:

• Aida Alvera-Azcárate (University of Liège, Belgium), Data analysis methods for oceanographic applications
• Satardu Bag (Max Planck Institute for Astrophysics, Garching), Discovering Strongly Lensed Transients in the Era of LSST: Machine Learning for a Needle-in-a-Haystack Problem
• Maria Benito (Instituto de Astrofísica de Canarias, Spain), Representation learning for dark matter searches in the Milky Way stellar halo

• Matteo Cagiada (University of Copenhagen and Oxford University), Towards improved prediction of protein dynamics

• Christopher Eckner (Instituto de Astrofísica de Canarias, Spain, ex SMASH Fellow), Application of Simulation Based Inference to astrophsyics
• Juergen Gall (Lamarr Institute for Machine Learning and Artificial Intelligence), From Forecasting Human Behavior to Forecasting Extreme Weather and Climate Events
• Imad El Haddad (Paul Scherrer, Switzerland), AI-Enabled Atmospheric Chemistry: Harmonized Measurements, Exposure Mapping, and Causal Inference
• Fabio Iocco (University of Naples, Italy), Determining dark matter distribution in Galaxies with machine learning
• Željko Ivezić (University of Washington, USA), Rubin Obs. science in the AI era, TBC
• Adriana Milic (CERN, Switzerland), The Art of Finding Needles in Haystacks: Event Selection at ATLAS
• Soebur Razzaque (University of Johannesburg, South Africa), Application of Machine Learning in Astrophysical Transient Data
• Johannes Schneider (University of Liechtenstein), How Using AI agents alters our thinking and speaking
• Gad Shaulsky (Baylor College of Medicine, Houston, USA), Deciphering Dictyostelium Biology: From Genetic Pathways to AI Systems for Knowledge and Discovery
• Dimitar Trajanov (University of Skoplje and Boston University), LLM based agentic systems and knowledge graphs
• Marta Via Gonzales (Barcelona Supercomputing Centre, ex SMASH Fellow), Machine-learning methods for aerosol source apportionment and harmonisation
• Saman Vinke, MD (Radboud University Medical Center), AI application to the field of Deep brain stimulation
• Lili Yang (Sun Yat-sen University, China), The application of ML in astrophysics and astroparticle experiments
• Jure Zupan (University Cincinnati, USA), Simulating Particle Physics Hadronization with Machine Learning

 

Monday 19 October

1 Data analysis methods for oceanographic applications

Speaker: Dr Aida Alvera-Azcárate (University of Liège, Belgium)

2 Discovering Strongly Lensed Transients in the Era of LSST: Machine Learning for a Needle-in-a-Haystack Problem

Speaker: Dr Satardu Bag ((Max Planck Institute for Astrophysics, Garching)

10:30 Coffee break

3 Machine Learning classification of Tidal Disruption Events from partial optical light curves

LSST is expected to find about several thousands of Tidal Disruption Events (TDEs) in its 10 years of operation. The physical mechanisms governing TDE emission, as well as the diversity of TDE subclasses, are only partially understood, and in order to clarify them, we require spectroscopic follow-up observations during the pre-peak rise phase and around peak brightness. This promotes attempts at developing early photometry-based detection of TDE candidates. In this talk, we investigate early classification strategies using the dataset developed for the MALLORN Kaggle challenge, which focused on nuclear transient classification with an emphasis on TDE identification. The dataset consists of LSST-like synthetic light curves and was originally designed for full light-curve classification. We repurposed this dataset to evaluate classification performance for incomplete light curves, in particular, at peak and pre-peak phases. We show that with partial light curves, the best F1 score drops from ~0.68 to 0.3-0.4, with over the half of the TDEs being missed during classification. The absence of post-peak color evolution and full event duration features strongly impacts model performance. We report possible ways of mitigate performance deterioration and what are the main contaminants to the TDEs, and consider the possibility that thorough follow-up photometric observations of transient events on the rise are needed (in addition to LSST), to provide us with the light curves with higher-cadence and finer wavelength coverage, so that we could search for better discriminating features for early TDE detection.

Speaker: Oleksandra Razim (University of Nova Gorica)

4 Anomaly Detection for Faint Galactic Substructure with EagleEye

Detecting faint stellar substructure in the Milky Way halo and its surroundings favours methods that remain sensitive to weak signals without restrictive assumptions about either the signal morphology or the Galactic background. We present recent applications of EagleEye, a model-independent anomaly detection framework that compares multidimensional data distributions to identify localized over- and underdensities relative to an empirical reference sample. We highlight two applications: the detection of faint dwarf galaxies around the Milky Way and the search for stellar wakes induced by its most massive satellites. As stellar wakes are expected to be exceptionally faint and difficult to model, yet offer a novel probe of dark matter substructure, they are a particularly exciting target for data-driven anomaly detection methods.

Speaker: Sven Põder (Scuola Internazionale Superiore di Studi Avanzati (SISSA))

5 Two-Phase Rapid Simulation-Based Inference with Differentiable Simulators

I will introduce introduce Rapid Simulation Based Inference (RSBI), a diffusionbased variational approach to likelihood-free Bayesian inference that achieves high sampling efficiency under large prior-to-posterior volumes with multi-modal posterior structure.
RSBI builds on advances in Schrödinger Bridge diffusion sampling to handle multimodal posteriors, with the likelihood initially supplied by a surrogate distance measure via a differentiable simulator, followed by neural ratio estimator trained on a constrained corpus of simulation/parameter pairs. The key innovation is that an appropriate surrogate likelihood gives a highly efficient proposal distribution for multimodal posteriors, replacing sequential rounds of other methods with a one-shot proposal. Subsequent NRE refinement targets the posterior under the simulator and original model prior.
As a variational method, RSBI does not suffer from leakage and implicit target drift commonly observed in standard sequential posterior estimation methods. To improve mode coverage we optionally utilize the well-tempered meta-dynamics framework, which also encourages exploration of the prior volume. We observe strong performance on standard SBI benchmarks, particularly when the prior volume is scaled up to 2500 times the original, achieving a performance degradation of only $\sim 5 \%$ on Two Moons under a highly constrained simulation budget. Additionally, we evaluate RSBI's performance for gravitational wave ring-down posterior estimation, in a real-world physics benchmark inspired by black hole spectroscopy.

Speaker: Andre Scaffidi (SISSA)

6 Transfer learning for regional data-driven weather prediction

Training high-resolution data-driven weather prediction models from scratch is computationally expensive and requires long regional datasets. We explore whether a pretrained regional machine-learning model can be transferred to a new geographic domain.
A graph neural network-based limited-area weather model trained over Switzerland is transferred to the Eastern Alps and Northern Adriatic in two ways: directly, without additional training, and through fine-tuning on five years of regional analysis data. This allows us to test both zero-shot generalization and target-domain adaptation across different grids, terrain, and climate conditions.
The pretrained model already produces meaningful forecasts when applied directly to the new domain. Fine-tuning further improves performance for most variables, while precipitation skill remains encouraging in both configurations.
The results show that transfer learning can reduce the data and computational requirements of data-driven regional weather prediction model development and offers a practical route for adapting existing machine-learning forecasting systems to new regions.

Speaker: Matjaž Puh (UNG, ARSO)

7 RiverTiles: An Open Web-Based Platform For Hosting And Visualizing Uhr River Imagery And Annotations

River monitoring increasingly relies on hyperspatial imagery capable of resolving geomorphological and ecological features at centimeter scales. Although UAVs and other low-altitude airborne platforms have made such data more accessible, resulting datasets often remain dispersed across individual projects, inconsistently documented, and rarely accompanied by reusable spatial annotations. This limits comparative analysis, reproducibility, and the development of transferable machine-learning methods for river environments. Here, we present RiverTiles, an open-access geospatial platform for the curation, visualization, and dissemination of ultra-high-resolution river imagery and hierarchical geomorphological annotations. RiverTiles combines a lightweight web architecture for interactive exploration of large raster datasets with a GUS-inspired annotation framework organized across macro-units, geomorphic units, and finer structural elements. The platform supports curated community contributions, standardized metadata, and publication of both visualization-ready and analysis-ready assets. RiverTiles brings imagery, terrain products, annotations, provenance information, and archival records into a common geospatial framework, providing shared infrastructure for reproducible river monitoring, geomorphological analysis, and future computer-vision workflows.

Speaker: Tulio Soto Parra (Faculty of Computer and Information Science, University of Ljubljana)

8 Learning Multi-Hazard Early Warning from Atmospheric State: Graph and State-Space Models for Alpine Catchments

Forecasting flash floods, landslides and debris flows in Alpine terrain is hard for a structural reason: the conventional warning chain runs from quantitative precipitation forecast (QPF) through hydrological model to alert, so its skill is capped by its least predictable link. Convective, orographically forced rainfall is routinely misplaced by numerical weather prediction (NWP). During the catastrophic August 2023 Slovenian floods, operational models captured the large-scale setup --- moisture transport, instability, orographic ascent --- yet failed to localise the extreme rainfall, limiting the spatial precision of catchment-scale warnings.

FLOODCAST, an early-stage project being developed with the Slovenian Environment Agency (ARSO), reframes the task as direct probabilistic classification of hazard occurrence from the atmospheric state. For catchment $c$, hazard type $h$ and lead time $\tau$, the model estimates
$$p\big(\,y^{h}_{c,\,t+\tau}=1 \;\big|\; \mathbf{s}_{c},\ \mathbf{a}_{c,t},\ \mathbf{X}_{\mathcal{G},\,t-L:t}\,\big),$$ where $\mathbf{s}{c}$ encodes static susceptibility, $\mathbf{a}{c,t}$ the antecedent land state (multi-layer soil moisture, snowmelt, antecedent precipitation), and $\mathbf{X}{\mathcal{G},\,t-L:t}$ the evolving NWP atmospheric state --- instability indices, integrated vapour transport, mid-tropospheric circulation, terrain-relative flow --- over a lookback window $L$ on a catchment graph $\mathcal{G}$. Forecast precipitation is downweighted rather than depended upon. The system couples static and dynamic layers. The static layer provides hazard-specific spatial priors $\mathbf{s}{c}$: the first, a machine-learned Slovenia-wide flood susceptibility map built from terrain, hydrographic, land-cover and soil predictors, conditions the dynamic hazard probabilities. For the dynamic layer, $\mathcal{G}$ comprises 227 catchment nodes with contiguity, directed-routing and synoptic edges; we are developing a spatio-temporal graph neural network with a state-space (Mamba) sequence backbone, whose linear-time scaling suits long antecedent windows, alongside gradient-boosted-tree and convolutional U-Net baselines, pretrained on ERA5 reanalysis and fine-tuned on operational forecasts.

Evaluation follows a leave-one-event-out protocol on historical Slovenian events, with AUPRC, critical success index and reliability diagrams as headline metrics, and probability calibration under severe class imbalance as a central design constraint. We will present preliminary susceptibility-mapping results and progress on an end-to-end pilot targeting the August 2023 event, discussing the transferability of environment-based hazard learning across domains.

Speaker: Aman Arora (University of Nova Gorica, Vipavska cesta 13, Rožna dolina, 5000 Nova Gorica)

12:30 Lunch

Welcome addresses + Round table on 'Shaping the Future of AI in Science: Funding Strategies and Opportunities'

15:30 Coffee break

9 Representation learning for dark matter searches in the Milky Way stellar halo

Speaker: Dr Maria Benito (Instituto de Astrofísica de Canarias, Spain)

10 Machine-learning methods for aerosol source apportionment and harmonisation

Speaker: Dr Marta Via Gonzales (Barcelona Supercomputing Centre, ex SMASH Fellow)

Tuesday 20 October

11 Simulating Particle Physics Hadronization with Machine Learning

Speaker: Dr Jure Zupan (University Cincinnati, USA)

12 AI-Enabled Atmospheric Chemistry: Harmonized Measurements, Exposure Mapping, and Causal Inference

Speaker: Dr Imad El Haddad (Paul Scherrer Institute, Switzerland)

10:30 Coffee break

13 Clustering for large-scale reinterpretation of new physics searches at the LHC

Beyond the Standard Model (BSM) searches show no statistically significant sign of new physics to date. However, several analyses reported small excesses, higher than 2σ SD beyond the SM expectation. In this work, clustering algorithms are used to extract additional insights from existing searches and to motivate a next round of BSM analyses. The flexible framework of the phenomenological Minimal Supersymmetric Standard Model (pMSSM) is used for a large-scale reinterpretation of new physics models using existing analyses. Models consistent with the observed excesses are explored using clustering methods. In this work, clustering algorithms are benchmarked for this high-dimensional problem. Optimal algorithms are deployed to perform a large-scale reinterpretation of new physics searches to help inform the direction of future searches at the HL-LHC.

Speaker: Judita Mamuzic (IJS, Ljubljana)

14 Anomaly detection with normalizing flows for model-agnostic new physics searches

Discovering new particles from beyond the Standard Model remains one of the main goals of present-day particle physics. Traditional searches for new physics at the Large Hadron Collider rely on specific theoretical scenarios and simulation-based background estimates, limiting their reach and introducing modeling uncertainties. We present an anomaly detection method that uses normalizing flows to identify anomalous events without assuming a particular signal model, with the goal of estimating backgrounds directly from data rather than simulation. By training a flow on data and splitting the resulting latent representation into two independent parts, we construct two decorrelated anomaly scores that allow the background in the signal region to be estimated using the well-established ABCD method. We test this approach on a benchmark new physics scenario and show that it successfully decorrelates the anomaly scores and correctly estimates the S/√B in the signal region.

Speaker: Rafał Masełek (Jozef Stefan Institute)

15 Integrating Photo-Thermal Aerosol Absorption Monitor and Single-Particle Spectroscopy with Machine Learning to Improve Mass Absorption Cross-Section Measurements of Black Carbon and Mineral Dust Aerosols

Light-absorbing aerosols, particularly black carbon (BC) and mineral dust (MD), remain among the most uncertain light-absorbing components in our understanding of Earth's radiative balance. The mass absorption cross-section (MAC) serves as the critical link between aerosol mass and light absorption, yet literature reported values vary dramatically –by a factor of four for BC alone (Wang et al., 2023). This variability stems not only from natural differences in particle sources and morphology but also from fundamental limitations in existing measurement approaches. In this study, we propose integrating two complementary techniques: photothermal aerosol absorption monitor (PTAAM) (Drinovec et al., 2022); and the single particle soot photometer (SP2). PTAAM measures aerosol light absorption with high sensitivity, while SP2 provides single-particle resolution of the scattering and incandescence signal of refractory particle, from which the rBC mass is derived (Tian et al., 2025). By coupling these instruments with advanced machine learning (ML) algorithms, particularly graph neural networks (GNNs), we aim to extract more physically meaningful MAC values from complex atmospheric datasets (Lamb et al., 2023). The methodology herein proposed involves several innovations. First, we will extend SP2 calibration beyond its traditional rBC focus to accurately quantify other refractory absorbing particles, such as iron oxide content in mineral dust. Second, multi-wavelength PTAAM measurements will be coupled with classifying instrumentation to achieve mass-resolved absorption data. Third, data fusion algorithms (i.e., GNN, Random Forest, XGBoost) will leverage the complementary strengths of both techniques. Transfer learning approaches, such as pretraining GNNs on laboratory or numerically generated aerosol populations and fine tuning them on field PTAAM–SP2 measurements, may help bridge the gap between controlled laboratory conditions and field observations, though the generalization of such models requires careful validation. Ultimately, this work will provide climate modelers with more constrained aerosol light-absorption parameters, potentially reducing uncertainties in radiative forcing estimates.

Speaker: Ankur Bhardwaj (Centre for Atmospheric Research, University of Nova Gorica)

16 Trustworthy species distribution modelling: reliability audits for inference from imperfect observations

Machine-learned models of species distributions are built, more often than not, on observations the modeller did not collect. Aggregated archives pool records from many contributors, instruments and eras, and with them a wide range of positional and observational quality. Our work asks one question across this pipeline: how far can the resulting predictions be trusted, and where exactly do they fail? It treats it as a problem of reliability propagation: uncertainty and contamination in the observations travel through the fitted model into the ecological conclusions drawn from it.
Occurrence records are routinely filtered by positional quality before modelling, a step almost universally treated as neutral hygiene. It is not. Filtering is neutral only if quality is unrelated to the predictors; when it is not, discarding low-quality records removes them non-randomly across predictor space, the distribution the model learns from shifts, and cleaning becomes a source of bias. We formalise this as a covariate-shift problem and develop a diagnostic that detects, without ground truth, whether and along which axes filtering displaces the data. Applied to a curated global freshwater-crayfish database (115,191 records) and to open-aggregator Odonata records, filtering shifts the observed distribution significantly and coherently; two defensible corrections move the model in opposite directions and bracket a truth that is not identifiable from the data; and stricter filtering induces more bias, not less.
The consequence does not stay small. Mean per-cell changes in predicted suitability of 0.03–0.05 become changes of 7–52% in predicted range area, the quantity conservation decisions actually use. Error is amplified, not absorbed, on the way downstream.
I will place this within a wider reliability layer under construction: auditing the observations that enter, calibrating the predictions that leave, where miscalibration proves spatially structured, concentrating in particular parts of the river network, and propagating the residual uncertainty into derived quantities, together with an emerging R implementation designed to sit alongside existing modelling stacks rather than replace them.
The lesson generalises beyond ecology: wherever a data-quality flag correlates with the predictors, cleaning is a modelling decision with consequences, not a hygiene step taken before the modelling begins.

Speaker: Kristian Miok (FRI UL)

17 Representation Learning for Coupled Earth-System State Estimation

Modern Earth-system models simulate interacting atmospheric, oceanic, land-surface, cryospheric, and chemical processes. Yet estimating the current state of these systems from incomplete and noisy observations remains a fundamental challenge. Current approaches rely on separate estimation procedures for data assimilation in different Earth-system components, leading to inconsistencies, suboptimal estimates, and adjustment shocks.
This project explores whether the relationships in the background-error covariance model, currently prescribed through analytical uncertainty models can instead be learned directly from data. To this end, we develop graph-neural-network autoencoders that compress the coupled Earth system into a common low-dimensional latent space, where observations and model predictions can be combined consistently.
The approach offers three potential advantages. First, it enables observations from one Earth-system component to influence others through learned cross-component relationships. Second, nonlinear latent representations may better capture the non-Gaussian behaviour of variables such as air humidity, sea ice thickness, salinity, or ozone. Third, dimensionality reduction in latent space improves computational efficiency while preserving physically meaningful structures. Beyond the methodological advances, more physically consistent state estimates can improve both weather forecasts and Earth-system reanalyses.

Speaker: Lukas Kugler (Faculty of Mathematics and Physics, University of Ljubljana)

18 AI for Security & Security for AI

Artificial Intelligence (AI) is reshaping cybersecurity by enhancing threat detection, automating incident response, and supporting intelligent security operations. At the same time, AI systems have become attractive targets for attacks such as adversarial manipulation, data poisoning, model extraction, and prompt injection. This dual role positions AI as both a powerful cybersecurity tool and a critical asset requiring robust protection.

This talk explores the intersection of AI for Security and Security for AI, highlighting the opportunities and challenges of deploying AI in adversarial environments. It discusses why conventional cybersecurity approaches are often insufficient for AI-enabled systems and presents emerging strategies for building resilient, trustworthy, and secure AI through security-by-design, adversarial robustness, and secure deployment practices. Drawing on recent research and real-world examples, the session provides practical insights and best practices for protecting AI-integrated systems while leveraging AI to strengthen modern cyber defense.

Speaker: Areeb Ahmed (University of Ljubljana)

AI for Security & Security for AI.pdf

12:30 Lunch

Presentation from Companies and Round table on transition between academia and industry

15:15 Coffee break

19 Diffusion-Generated Synthetic Ear Images for Cross-Dataset Recognition

Synthetic biometric data is attractive when privacy constraints, annotation cost, and limited subject availability restrict the collection of large real datasets. This paper presents a diffusion-first synthetic ear generation and benchmarking framework. The proposed pipeline uses Stable Diffusion 2.1 image-to-image synthesis: real ear crops define the source identity structure, multiple candidates are sampled per planned variant, candidates are scored by an ear-recognition encoder, and accepted images are filtered by identity consistency, visual quality, duplicate risk, and privacy diagnostics before verification-oriented adaptation. The completed diffusion run generated 1,048 candidates from 131 source identities and retained 169 images from 90 identities. Under the fixed template-quality protocol, base/adapted score fusion obtains 17.89\% EER, 6.09\% template EER, and 52.74\% Rank-1 on \earvn; 7.34\%, 0.69\%, and 93.47\% on AMI; 4.02\%, 2.76\%, and 93.83\% on IITDelhi; and 33.53\%, 17.88\%, and 7.00\% on UERC. Against the strongest listed non-ours synthetic baseline per target, the proposed diffusion method reduces mean EER from 25.92\% to 15.69\%, a 39.5\% relative reduction. A few-real ablation shows that, with the accepted diffusion set fixed, adding two real images per identity gives the best mean EER (15.61\%) and eight real images per identity gives the best mean Rank-1 (61.76\%).

Speaker: iyyakutti ganapathi

20 Heart failure prediction using non-invasive voice biomarkers using machine learning and SHapley Additive exPlanations

Heart failure (HF) is a chronic disease all around the world, affecting millions of people. Despite recent advances in the diagnosis of heart failure, the traditional diagnosis is highly invasive and time-consuming. The disease results in physiological changes in patients, including increased vocal edema in the vocal tract and accumulation of fluid in the lungs, thereby affecting the laryngeal and respiratory systems. These physiological changes ultimately affect the voice patterns that can be detected using artificial intelligence techniques. Moreover, heart failure patients exhibit more pauses during speech, with increased voice hoarseness due to congestion and early exhaustion during physical activity, including multiple voice-based activities. Here, we present a non-invasive approach to predict heart failure patients using voice biomarkers and features that capture changes in voice patterns compared with those of suspected heart failure patients. We compare and evaluate the performance of using five voice recordings per patient, each containing 94 extracted features, for a total of 470 features. The high-dimensional feature space is reduced using SHapley Additive exPlanations (SHAP) to select important features, and the results are compared with those from manually selected feature groups. Moreover, using top-k SHAP-based feature selection as a hyperparameter during cross-validation with varying k across folds, thereby affecting the final reported features’ frequency and stability, is also reported and compared with the fixed, appropriate top-k selection.

Speaker: Muniba Ashfaq (University of Ljubljana, Faculty of Computer and Information Science)

21 AI application to the field of Deep brain stimulation

Speaker: Saman Vinke, MD (Radboud University Medical Center, Netherlands)

Round table on Future of Science and Academia - open to the public

Wednesday 21 October

22 How Using AI agents alters our thinking and speaking

Speaker: Dr Johannes Schneider (University of Liechtenstein)

23 Towards improved prediction of protein dynamics

Speaker: Dr Matteo Cagiada (University of Copenhagen and Oxford University)

10:30 Coffee break

24 Toward the first syntactically annotated corpus of Slovenian dialects: an LLM-assisted pipeline from raw transcripts to Universal Dependencies

Slovenian dialects — among the most diverse in the Slavic world — remain almost invisible to language technology: no dialect corpus with morphosyntactic annotation exists, and standard NLP tools fail on dialect input. I present work toward the first Universal Dependencies (UD) annotated corpus of Slovenian dialects, built with a modular, LLM-assisted pipeline.
The pipeline has four stages. (1) Collection: merge and quality-filter heterogeneous sources (e.g., GOKO, Fran dialect dictionaries, and others smaller sources) and mine dialect–standard glossaries from inline glosses, dictionaries, and word-aligned transcriptions. (2) Normalization: dialect forms are mapped to standard Slovenian through a cascade of deterministic phonological rules, lexicon lookups expanded via Sloleks inflectional paradigms, and — only for unresolved tokens — the Slovenian generative model GaMS as a constrained word-level oracle, under a strict one-to-one token mapping that preserves alignment with the original forms. (3) Annotation: normalized text is parsed with CLASSLA-Stanza; dialect forms are retained and each token receives a variety tag (standard/non-standard/dialectal). (4) Gold data and fine-tuning: a manually corrected subcorpus feeds fine-tuning of SloBERTa and GaMS, closing the loop between rule-based and neural components.
I report first results on normalization accuracy and parser performance, discuss possible synthetic dialect data generation as an augmentation strategy, and argue that hybrid rule–LLM cascades are a realistic path to treebanks for non-standard varieties of low-resource languages.

Speaker: Varvara Magomedova (UNG)

25 Structural plausibility without binding specificity: what confidence scores do and don't tell us about antibody–antigen prediction

Antibody discovery increasingly uses AI structure prediction to select candidate binders before experimental testing. Such selection assumes that a model's confidence score reports on binding specificity, and not merely on the plausibility of the predicted complex.
We benchmarked AlphaFold3, Boltz-2 and Chai-1 on 106 nanobody–antigen and 46 antibody–antigen complexes. Each binder was paired with every antigen (cognate and non-cognate matrices), with 50 predictions per pair (over 15,000 per model). Predictions were scored for confidence (ipTM) and for accuracy against the experimental structure (DockQ).
Cognate pairs were generally well predicted, but separation from non-cognate pairs was incomplete (ROC-AUC 0.77 to 0.87). Miscalibration was model-specific: Boltz-2 was systematically overconfident, Chai-1 underconfident. Repeated sampling within a single seed improved DockQ by 0.2 to 0.3, while confidence remained largely unchanged (variation 0.04 to 0.1), indicating that the models do not register an improved pose. Most trajectories reached a plateau after 10 to 25 samples. Oracle selection of the best of 50 predictions outperforms selection by confidence, but requires knowledge of the correct answer, which is not available in a screening setting.
Confidence metrics describe structural plausibility rather than binding specificity, and are not yet suitable as proxies for affinity.

Speaker: Klara Kropivšek Brumat (University of Nova Gorica)

26 A pipeline for shortlisting de novo designed protein binders towards CDKL5

Artificial Intelligence (AI) is now considered a mature technology for the de novo design of protein binders, offering a potentially more reliable option when working with poorly immunogenic or highly conserved target antigens, or when a specific conformational epitope must be targeted. Current benchmarks indicate that more than 10 designs typically must be characterized to yield binders with low nanomolar affinity and sufficient stability. However, most published successes have focused on biomarkers (e.g., PD-L1) for which structural data and targeting interfaces were already established via classical approaches. In contrast, there are very few documented cases where novel biomarkers and functionally relevant epitopes have been successfully targeted.
CDKL5 is an essential kinase whose dysfunction correlates with neurological conditions, including CDKL5 Deficiency Disorder. The sequence of its N-terminal kinase domain is highly conserved among CDKL enzymes, whereas its C-terminus is disordered and contains several evolutionary conserved sequences likely involved in functional regulation. These structural characteristics explain the historical difficulty in recovering highly specific antibodies for CDKL5 epitopes using classical discovery techniques. Consequently, the ability to quantify and precisely localize the protein in cells and tissues has been severely impaired.
Our project devised a customized approach to score de novo redesigned protein mini-binders targeting a specific, dynamically stable conformational epitope unique to the catalytic domain of CDKL5. The protocol integrated AI-driven biophysical models and physics-based simulations to shortlist the candidate pool to 15. The preliminary results will be discussed. Furthermore, we will show how a similar approach can be used to target also specific functional linear epitopes.

Speaker: Marco Orlando (University of Nova Gorica)

27 Mito-Syn: Mitosis Synthesis via Chromatin Modelling and Language-Grounded Inpainting

Detecting and subtyping mitotic figures in histopathology is central to tumor grading as well as patient survival prognosis, yet training data for rare phases and atypical mitoses remains scarce and costly to annotate, thus, generating synthetic data for these cases is desirable. Generative models have been introduced in the past that copy and paste mitotic chromatin structure from one context to another, yet the augmentation capability of such models are limited only to the context and not the chromatin structure itself.

We introduce Mito-Syn, a two-stage pipeline that decouples chromatin structure (location and shape) from tissue texture, grounding each stage in a different form of domain knowledge. First, a conditional generative shape model, conditioned on mitotic phase and cell boundary geometry, trained on thousands of mitotic cells from different mitotic subphases, synthesizes a binary chromatin mask. Second, a conditional inpainting model paints the tissue texture inside the cell boundary, conditioned on mitotic phase and the chromatin mask generated by stage 1. For the second stage, we also introduce an alternative language-grounded inpainting technique, where the inpainting is conditioned on captions made by a pathology domain expert, focusing on the distinct characteristics of the specific mitotic phases.

We showcase the augmentation value of our method on publicly available mitosis subtyping benchmark datasets. Our pipeline yields controllable, biologically consistent synthetic mitotic figures across the full phase spectrum. By explicitly disentangling morphology from appearance and incorporating expert language descriptions as a conditioning signal, our approach offers an interpretable and controllable alternative to end-to-end image generation for medical image synthesis.

Speaker: Dominik Hirling (University of Ljubljana, Faculty of Computer and Information Science)

28 From Retrieval to Trustworthy Answers: Practical Lessons from Building RAG Systems for Scientific Evidence

Retrieval-augmented generation (RAG) is becoming an important tool for scientific literature search and evidence synthesis. Here I introduce the core ideas behind RAG through practical examples from the BioASQ and TREC-RAG benchmarks. The presentation covers evidence retrieval, reranking, grounded answer generation, and evaluation, together with common pitfalls and practical lessons learned from developing competitive systems. The presentation is intended for researchers and fellows from different backgrounds who are interested in using RAG tools in scientific research and understanding its current capabilities and limitations.

Speaker: Yun Wang (University of Ljubljana)

29 Fermipylot: An AI Assistant for Automated Fermi-LAT Data Analysis

The Large Area Telescope (LAT) onboard NASA’s Fermi Gamma-ray Space Telescope continuously surveys the high-energy gamma-ray sky, providing observations of astrophysical phenomena ranging from variable sources and transient events to large-scale diffuse emission. Its extensive public data archive can be analysed using Fermipy, an open-source Python package that provides a high-level interface to the Fermi Science Tools and supports standard gamma-ray analysis workflows. Together, these public resources offer an ideal opportunity to explore how agentic AI systems can automate and manage complex scientific workflows.

We present the concept of Fermipylot, an AI assistant designed to guide and automate Fermi-LAT data analysis. The idea is to combine a Large Language Model (LLM) with an agentic retrieval-augmented generation workflow, grounding its decisions in Fermipy technical documentation, source code, and tutorial notebooks. From a request expressed in Natural Language (NL), Fermipylot will be able to understand the scientific goal, retrieve the relevant information, and prepare and execute a reproducible analysis. The agent will evaluate the correctness of the retrieved context and check its generated responses, while deterministic components control the execution of the scientific workflow. Assumptions, configurations, intermediate products, and results will be recorded to make each analysis transparent and reproducible.

Fermipylot will provide a case study of how LLM-based agents can connect NL interaction with established scientific software. More broadly, it will explore how AI can facilitate access to specialised research tools and support the preservation, sharing, and transfer of scientific expertise across communities and generations of researchers.

Speaker: Ermes Aviano (University & INFN Trieste)

30 Resource-Efficient LLMs for Depression Symptoms Screening: Performance and Limitations in Zero Shot Setting

Depression is the leading cause of global disability and early detection is crucial for effective intervention. Recent advances in large language models (LLMs) offer potential for analyzing text to identify depression symptoms. This work investigates the zero-shot capability of LLMs to recognize nine DSM5 depression symptoms from short-text
inputs. We evaluated eight open LLMs with model sizes ranging from 1.5B to 14B parameters using a clinically annotated dataset and assessed both overall agreement and symptom-level performance. Results indicate that
while smaller models exhibit limited clinical accuracy, the Qwen 2.5-7B model achieves substantial performance with a Cohen’s Kappa of 0.603 and a Macro F1 score of 0.648. Notably, a performance plateau between the 7B and 14B Qwen variants suggests that model scaling alone does not guarantee improved symptom-level classification, establishing Qwen 2.5-7B as a resource-efficient model. Further analysis of the best-performing model revealed strengths in identifying salient symptoms like suicidal thoughts, but limitations in recognizing core symptoms such as
depressed mood and anhedonia. Misclassification analysis reveals that the model frequently misclassifies posts expressing ’depressed mood’ as ’no symptom’ or vice versa, often overlooking indicators of irritability or social withdrawal. These findings suggest that resource-efficient LLMs can support preliminary symptom screening in zero
shot settings, but there is risk of overlooking clinically important symptoms without fine-tuning.

Speaker: Muhammad Rizwan (University of Ljubljana)

12:45 Lunch

31 From Forecasting Human Behavior to Forecasting Extreme Weather and Climate Events

Speaker: Dr Juergen Gall (Lamarr Institute for Machine Learning and Artificial Intelligence, Germany)

32 Application of Machine Learning in Astrophysical Transient Data

Speaker: Dr Soebur Razzaque (University of Johannesburg, South Africa)

33 Determining dark matter distribution in Galaxies with machine learning

Speaker: Dr Fabio Iocco (University of Naples, Italy)

16:00 Coffee break

Public event - scientific afternoon

Thursday 22 October

34 Application of ML in astrophysics and astroparticle experiments

Speaker: Dr Lili Yang (Sun Yat-sen University, China)

35 Deciphering Dictyostelium Biology: From Genetic Pathways to AI Systems for Knowledge and Discovery

Speaker: Dr Gad Shaulsky (Baylor College of Medicine, Houston, USA)

36 Application of Simulation Based Inference to astrophysics

Speaker: Dr Christopher Eckner (Instituto de Astrofísica de Canarias, Spain, ex SMASH Fellow)

11:00 'XL' Coffee break

37 Luminosity function of millisecond pulsars in globular clusters from their gamma-ray spectra

Millisecond pulsars (MSPs) are widely believed to be responsible for the gamma-ray emission of globular clusters (GCs), yet the underlying MSP gamma-ray luminosity function remains uncertain. Existing GC-based determinations of the latter often rely on external prescriptions for the number of MSPs in each cluster.

In this work, we constrain the MSP gamma-ray luminosity function in Milky-Way GCs directly from gamma-ray spectral energy distributions (SEDs), establishing an inference strategy that remains computationally viable for joint analyses of large GC samples.
We construct a forward model in which each GC hosts a population of MSPs whose luminosities are drawn from a universal log-normal luminosity function. Synthetic GC SEDs are generated by summing individual MSP spectra, and we infer posterior distributions using first a likelihood-based nested sampling and then a Bayesian implicit-likelihood inference in the neural ratio estimation (NRE) approach.

We show how the likelihood-based nested sampling approach becomes rapidly impractical as the number of jointly modelled clusters (and cluster-specific latent parameters) increases, whereas the NRE approach scales mildly with sample size and reproduces the likelihood-based posteriors in regimes where direct sampling is feasible.

We then apply the implicit-likelihood pipeline to Fermi-LAT SEDs from the 4FGL-DR4 catalog for a sample of 36 GC-associated sources, augmented by dedicated phase-resolved ON/OFF SEDs for two clusters hosting individually detected MSPs.

For the real data sample, we obtain informative constraints on the luminosity-function parameters consistent with previous analyses that constrain the cluster-specific MSP counts via complementary multi-wavelength information.

Speaker: Francesco Xotta (University of Nova Gorica)

38 Simulation-based cosmological inference from 3D maps of multiple tracers

Extracting maximum cosmological information from current and upcoming large-scale structure data requires going beyond summary statistics as currently used in likelihood-based inference. Simulation-Based Inference (SBI) promises to enable the exploitation of field-level information and the rich physics of modern hydrodynamical simulations. We develop a proof-of-concept SBI pipeline to explore its potential to constrain the cosmological parameters $\{\Omega_m,\sigma_8\}$ from galaxy number counts, neutral hydrogen (HI) intensity mapping and their combination. We use neural emulators trained on full hydrodynamical simulations to generate galaxy and HI maps from fast, approximate dark matter simulations. Combined with neural posterior estimation, this enables the estimation of cosmological parameters while marginalizing over astrophysical effects. We perform inference both on the power spectrum and on representations derived from field-level 2D or 3D maps, comparing results from each probe and the combination of both tracers, and assessing the impact of data compression and multi-tracers information on cosmological constraints. Combining galaxy and HI fields improves constraints with respect to single-tracer cases by a factor 2 to 7 in terms of a Figure of Merit describing the joint precision on cosmological parameters, depending on the tracer/configuration. Moving from summary statistics to field-level inference leads to a consistent gain in constraining power of about a factor 3, with 3D maps providing the most precise and well-calibrated posteriors. This gain in precision is robust even when astrophysical parameters are marginalized over. Further developments (including realistic survey effects and improvements in emulators' faithfulness) will enable the application of this analysis pipeline to upcoming surveys.

Speaker: Giulio Scelfo (SISSA, INFN)

39 Simulation-Based inference for massive black hole binary from mock LISA data

The Laser Interferometer Space Antenna (LISA) will observe gravitational waves produced by several massive black hole binary (MBHB) mergers per year. While the likelihood can be written in closed form under idealised stationary, Gaussian noise assumptions, including realistic effects — instrumental glitches, gaps in the data, and non-stationary noise — make it intractable or computationally prohibitive. This motivates the development of Simulation-based inference (SBI) pipeline, which requires only forward simulations of the data and can thus in principle include arbitrary complex physics. We present proof-of-concept results using truncated marginal neural ratio estimation (TMNRE), a sequential SBI method that iteratively truncates the prior to the region supporting non-negligible posterior mass. One of the challenges is the enormous shrinkage of prior volume under the posterior, a factor $\sim 10^{-24}$, which makes ordinary MCMC and sequential SBI difficult. We demonstrate inference of the parameters of a single MBHB from frequency-domain LISA data with a customised TMNRE algorithm, accommodating the periodicity of the angular parameters and preserving multimodal support during truncation, and validate the method against the MCMC posterior of the tractable case (obtained by artificially reducing the prior volume around the fiducial parameters). For a mock observation with chirp mass $10^{5.25} M_\odot$ and SNR 2200, we correctly recover the MCMC posterior over the 11-dimensional parameter space, in about 30 hours of training on a NVIDIA A100 GPU.

Speaker: Gianmarco Puleo (Scuola Internazionale Superiore di Studi Avanzati (SISSA))

40 Rapid Detection and Inference of Extreme-Mass-Ratio Inspirals in LISA: Divide and Conquer

Extreme-mass-ratio inspirals (EMRIs) are key gravitational-wave sources for the Laser Interferometer Space Antenna (LISA), but their detection and parameter inference are computationally challenging due to the extreme concentration of posterior distributions within vast prior volumes. In this work, we introduce a novel divide-and-conquer strategy that reformulates global inference as a hierarchical identification problem. Our approach iteratively localizes the posterior mode through a coarse-to-fine procedure based on ordinal classification, progressively restricting the parameter space while preserving the true signal region. A transformer-based neural network is trained at each stage to identify the most probable parameter subregions, enabling exponential reduction of the search volume with only a few refinement steps. Once the parameter space is reduced to the Fisher-information scale, standard local sampling methods efficiently recover the full joint posterior. We demonstrate that this method achieves rapid and accurate intrinsic parameter estimation for EMRIs in simulated LISA data, dramatically reducing computational costs compared to traditional global sampling techniques. This framework provides a scalable and efficient pathway for real-time EMRI detection and inference in the LISA era.

Speaker: Rahul Srinivasan (SISSA, Italy)

41 DeepExtractor: Model-Agnostic Signal and Glitch Reconstruction for Gravitational-Wave Astronomy

Speaker: Tom Dooney (NIKEF)

12:45 Excursion, wine tasting and conference dinner

Friday 23 October

42 Rubin Obs. science in the AI era

Speaker: Dr Željko Ivezić (University of Washington, USA)

43 LLM based agentic systems and knowledge graphs

Speaker: Dr Dimitar Trajanov (University of Skoplje and Boston University)

44 The Art of Finding Needles in Haystacks: Event Selection at ATLAS

Speaker: Dr Adriana Milic (CERN, Switzerland)

11:00 Coffee break

45 MicroSight: Low-Cost, Image-Based Identification and Classification of Microplastics in Freshwater Environments

Microplastic (MP) pollution in freshwater systems remains poorly monitored due to the cost and complexity of existing identification methods, which typically rely on high-magnification imaging or laboratory-based spectroscopy. MicroSight addresses this gap by developing a machine learning pipeline for automated MP classification from images captured with consumer-grade devices, enabling scalable, field-deployable monitoring.

The project's first phase focuses on building an open, annotated image dataset of MP particles (fragments, fibers, films, foams, pellets) prepared under controlled laboratory conditions and varied environmental backgrounds, including sediment and water matrices, biofouling, and weathering, to reflect real-world sampling scenarios. Particle segmentation and annotation combine pre-trained foundation models with manual verification, and physically meaningful shape descriptors (aspect ratio, roundness, solidity, convexity) are extracted to support morphology-based classification alongside deep learning features.

We will present the first labelled iteration of this dataset, along with baseline classification results from an initial ML model, including an early assessment of how image resolution affect classification performance. These results lay the groundwork for an open-access platform for crowdsourced MP recognition, offering a reproducible, low-cost framework for large-scale environmental monitoring and a foundation dataset for the broader scientific community.

Speaker: Nerea Portillo De Arbeloa

46 Statistical Approaches for Dark Matter Signatures in Astrophysical Neutrino Data

High-energy astrophysical neutrinos offer a unique opportunity to investigate possible interactions between dark matter and Standard Model particles. In particular, neutrinos produced near active galactic nuclei may traverse the enhanced dark matter densities expected around supermassive black holes, where dark matter--neutrino scattering could produce observable attenuation and spectral redistribution in the detected neutrino flux. We develop a likelihood-based statistical framework to search for such signatures using publicly available IceCube point-source data from four candidate neutrino-emitting active galaxies: NGC~1068, NGC~4151, TXS~0506+056, and PKS~1424+240. The expected event distributions are calculated by incorporating dark matter density-spike models, neutrino production spectra, detector effective areas, and propagation effects. Both energy-independent and linearly energy-dependent dark matter--neutrino scattering cross sections are considered. Constraints are obtained using a Poisson likelihood-based $\chi^{2}$ analysis, first for each source independently and subsequently through a joint stacking analysis in which a common dark matter mass and interaction cross section are fitted across all four sources. The combined analysis significantly improves the sensitivity relative to individual-source searches and yields tight upper limits on the dark matter-neutrino scattering cross section for an energy-independent interaction under the most optimistic dark matter spike scenario. The statistical importance of individual sources depends on both their event statistics and their energy distributions. The results are also interpreted within an anomaly-free $U(1)_{L_{\mu}-L_{\tau}}$ model containing pseudo-Dirac or complex-scalar dark matter, demonstrating the potential of statistically combined astrophysical-neutrino observations to constrain otherwise inaccessible dark-sector parameter space.

Speaker: Khushboo Dixit (Centre for Astro-Particle Physics, University of Johannesburg)

47 Probing the Parameter Space of Axion-Like Particles Using Simulation-Based Inference

Axion-like particles (ALPs) appear in various extensions of the Standard Model and can interact with photons, leading to ALP-photon conversions in external magnetic fields. This phenomenon can introduce characteristic energy-dependent “wiggles” in gamma-ray spectra. The Cherenkov Telescope Array Observatory (CTAO) is the next-generation ground-based gamma-ray observatory, designed to provide enhanced sensitivity and energy coverage (20 GeV – 300 TeV) over current Imaging Atmospheric Cherenkov Telescopes (IACTs) and offers an excellent opportunity to study such effects.

In this work, we employ Simulation-Based Inference (SBI) to explore the parameter space of ALPs, targeting the flaring states of blazars, which are among the brightest gamma-ray sources and ideal candidates for probing ALP-induced spectral modulations. Additionally, we investigate whether this inference method can reproduce ALP exclusion limits comparable to those reported in previous studies using the classical likelihood-ratio approach. This study therefore provides an assessment of SBI as a tool for constraining ALP–photon interactions.

Speaker: Pooja Bhattacharjee (University of Nova Gorica)

48 High-energy Multi-messenger Emission from Galaxy Clusters in the Local Universe

In this work, we investigate the diffuse $\gamma-$ray and neutrino emission from the Virgo, Perseus, and Coma clusters using a detailed numerical method, combining MHD simulations with Monte Carlo methods. The MHD simulation provides the distributions of temperature, gas, and magnetic field in clusters. The Monte Carlo simulations are used to investigate the cosmic-ray (CR) propagation in ICM and subsequently the secondaries stemming from CRs. Our primary assumption is that CR injection scales with the gas density of clusters, providing a physically motivated approximation. High-density regions in clusters are associated with strong turbulence and prominent shock structures, making them natural sites for efficient CR acceleration. Our predicted $\gamma-$ray flux from the individual clusters lies well below the present LHAASO upper limits. The MAGIC observations of the central source NGC~$1275$ of the Perseus cluster are significantly higher than our results. Further, we estimated the cumulative $\gamma$-ray and neutrino fluxes from clusters with masses $\gtrsim 5\times 10^{13}\, M_{\odot}$ in the local Universe (within $500$~Mpc). The diffuse $\gamma-$ray flux reported by the Fermi-LAT collaboration is significantly higher than our results. Our predictions are consistent with IceCube’s existing upper limits on the unresolved neutrino flux from galaxy clusters ($M > 10^{14}\, M_{\odot}$) up to $z = 2$.

Speaker: Saqib Hussain (Center for Astrophysics and Cosmology, University of Nova Gorica, Slovenia)

Clossing of the workshop and steps ahaed for the SMASH network

13:00 Lunch