HAMLET-PHYSICS 2026 Conference

19 Aug 2026, 06:00 → 21 Aug 2026, 20:00 Europe/Copenhagen

Margrethe Bohr Salen (Niels Bohr Building)

We are looking forward to welcoming you to the third annual HAMLET-PHYSICS Conference, to be held in sunny Copenhagen, August 19 - 21, 2026. It follows two very successful editions in 2024 and 2025. The workshop has three main goals: 1. To bring together Danish and international physicists using ML to meet, share ideas, and build community across location and physics specialty 2. To bring domain scientists into close contact with ML experts, to build community across the theory - application bridge 3. To provide a friendly environment for researchers to share best practices, for students to interact with experts, and for other sciences and industry to understand the state of ML in physics

Scientific Program

• Keynotes, plenaries and parallels
• Discussions and (AI-assisted) research speed-dating
• Beer talks and train-ride chats
• Hackathons and demonstrations from experts in high performance computing and machine learning

 

Abstracts are open for contributions at the intersection of machine learning and

• Particle physics
• Astrophysics and cosmology
• Quantum physics
• Biophysics
• Climate science
• Geophysics
• Molecular physics
• Condensed matter

This is not an exhaustive list. We warmly welcome all submissions for talks, suggestions and ideas, and will strive to accommodate all submissions.

 

Keynote Speakers

	Daniel Whiteson University of California, Irvine Daniel Whiteson is a Professor of Physics & Astronomy at UC Irvine, where he also holds a joint appointment in Logic and Philosophy of Science. He was one of the earliest researchers to bring deep learning into experimental physics, and his work on inference and learning ranges widely — from the ATLAS experiment at CERN to CRAYFIS, a distributed project that turns ordinary smartphone cameras into a global cosmic-ray detector, and machine-learned exploration of the space of possible universes. He is also a leading science communicator, co-authoring popular-science books with cartoonist Jorge Cham and co-hosting a widely followed physics podcast.
	Cecilia Clementi Free University of Berlin Cecilia Clementi is an Einstein Professor of Physics and Professor of Theoretical and Computational Biophysics at Freie Universität Berlin. Her work brings together statistical physics, molecular simulation and machine learning to understand biomolecular processes at long timescales, from protein folding to conformational change and molecular function. She has helped define the modern interface between ML and molecular simulation, including work on machine-learned coarse-grained models such as CGnets and CGSchNet, which use neural networks to simulate protein dynamics far more efficiently while retaining molecular detail. She is an ELLIS Fellow in the Machine Learning for Molecule Discovery programme, connecting her work to the broader European AI-for-science community.
	Shah Rukh Qasim ETH Zurich Shah Rukh Qasim works at the interface of machine learning and reconstruction. He is a co-developer of object condensation — a general, one-shot technique for reconstructing many objects at once from graph- and image-structured data — and of GravNet-style distance-weighted graph neural networks. Much of his work applies these methods to reconstructing particles directly from detector hits in high-granularity calorimeters such as the CMS HGCAL, and extends to real-time graph-network inference on FPGAs.
	Gabriel Facini University College London Gabriel Facini is an Associate Professor of Data Intensive Science and Physics at UCL and a core member of its Centre for Data Intensive Science and Industry, which brings machine learning to bear across astrophysics, high-energy physics, computer science and industry. He co-organises UCL's interdisciplinary AI for Data-Intensive Science and Industry programme and supervises students on applied ML placements with partners across media, analytics and the public sector. Within fundamental physics he works on charged-particle tracking, jet reconstruction and flavour tagging, and has helped carry modern architectures — such as transformer- and MaskFormer-based models drawn from computer vision — into scientific reconstruction.

Important Dates

• Registration & abstract submission opens: April 13, 2026
• Abstract deadline: July 3, 2026
• Notification of talks: July 10, 2026
• Program online: August 1, 2026
• Registration deadline: August 1, 2026
• Scientific program of the conference begins August 19, 09.00
• Scientific program of the conference ends August 21, 17.00

Abstracts submitted after the deadline will be considered on a case-by-case basis.

Social Program

Wednesday August 19th will feature a poster session and reception event at the University of Copenhagen Biocenter.

On Thursday evening August 20th, the workshop will take to the rails: A heritage 1950s Norwegian State Rail diesel locomotive will take workshop attendees from Copenhagen (Østerport Station) to Kronborg Castle in Helsingør (location of Shakespeare's Hamlet tale).

While on board, refreshments will be served, and breakout sessions will occur according to attendees research areas of interest. A visit of Kronborg will be included, along with an open-air conference dinner in Helsingør. 

Organization

Local Organizing Committee

• Daniel Murnane (NBI)
• Troels Petersen (NBI)
• Inar Timiryasov (NBI)
• Jean-Loup Tastet (DIKU)
• Troels Haugbølle (NBI)
• Oswin Krause (DIKU)

 

Wednesday 19 August

08:30 Registration and Coffee

1 Welcome

2 Keynote 1: Daniel Whiteson

10:25 Coffee

Plenary: Fundamental physics, inference, theory

3 Speeding up Analytic Calculations in Theoretical Particle Physics via Machine Learning

In this talk, we will present recent progress on applying machine-learning techniques to speed up calculations in theoretical physics, in which we desire exact and analytic results. One example are so-called integration-by-parts reductions of Feynman integrals, which pose a frequent bottleneck in state-of-the-art calculations in theoretical particle and gravitational-wave physics. These reductions rely on heuristic approaches for selecting a finite set of linear equations to solve, and the quality of the heuristics heavily influences the performance. In this talk, we present how a variety of machine-learning techniques (RL, evolutionary strategies and language agents) can identify improved heuristics that speed up the reductions by several orders of magnitude.

Speaker: Dr Matthias Wilhelm (NBI)

4 Normalising flows for all-orders QED corrections in lattice field theory

Many low-energy precision tests of hadronic observables in the Standard Model now require the inclusion of isospin-breaking corrections due to light-quark mass differences and electromagnetism (QED). Isospin breaking typically appears for any prediction with better than percent-level precision, where examples are particle decay rates, hadron mass differences and the muon anomalous magnetic moment. Lattice quantum field theory is a first-principles approach to calculate hadronic observables in a systematically improvable fashion. In this talk I will present how normalising flows and machine learning can be used to calculate all-orders electromagnetic corrections in lattice field theory, generally bypassing the complexity in Wick-contraction diagrams needed at fixed order. The new method is applied to lattice scalar QED in two, three, and four spacetime dimensions, yielding estimates with significantly reduced variance with respect to standard methods. Flows can be trained using small lattice geometries and subsequently evaluated on much larger lattice geometries while maintaining good efficiency. A generalisation to theories with fermions is envisaged, suggesting a path to applications in challenging field theories including lattice QCD. The talk is based on [hep-lat/2605.22444].

Speaker: Nils Hermansson-Truedsson (University of Edinburgh)

5 Model-agnostic searches for new physics with NPLM, and its reinterpretation

With recent advances in NSBI techniques, specifically neural ratio estimation, Beyond the Standard Model searches can now exploit higher dimensional inputs in an unbinned way, boosting their sensitivity for both parameter estimation and new physics searches. The New Physics Learning Machine (NPLM) approach (D’Agnolo et al., 2305.10500) provides a framework to adapt these techniques into an optimally sensitive anomaly detection algorithm. In this talk, I will outline the NPLM approach, focusing on the practical implementation required for analysis, before demonstrating a novel two-sample-test procedure that provides model-dependent parameter estimation.

Speaker: Jack Harrison (IFAE - Barcelona)

6 Hyperparameter Transfer: A Recipe for Efficient and Robust Scaling

The performance of deep learning models is sensitive to the settings of various hyperparameters, most importantly learning rate and initialization scale. Hyperparameter tuning becomes expensive at scale, motivating the modern paradigm of hyperparameter transfer, a strategy for inferring near-optimal hyperparameters in larger models from their smaller counterparts, which are cheaper to optimize. Though hyperparameter transfer is increasingly common in, e.g., large language models trained on textual data, applications to science remain relatively sparse. Furthermore, little progress has been made towards hyperparameter transfer in graph-based algorithms, which are particularly important for modeling relational data common in scientific datasets. We present systematic prescriptions for hyperparameter transfer in Graph Neural Networks (GNNs) and set-based models generalized to higher-order interactions, which we dub Coupled Particle-Edge Networks (CPENs). We derive and empirically validate learning rate transfer parameterizations for training via SGD, Adam, and AdamW, with applications spanning key benchmarks and scientific datasets including jet identification in particle physics and stellar stream identification in astronomical datasets. Our results indicate that proxy-tuning a small model leads to robust performance at scale, paving the way for applying large GNNs and CPENs to scientific data.

Speaker: Gage DeZoort (Princeton University)

7 Unfolding - Quantification learning for physics applications

Any physics experiment consists, in essence, of collecting times, locations and charges. The quantity of interest that the experiment measures must be derived from these three basic ones, for example via a neural network that takes these data as input and produces an estimate of the value in concern as output. However, the true distribution of the quantity in interest is convolved (or folded) with the response of the detector and the calculation method used in the experiment. Unfolding refers to the technique of recovering the true distribution of a quantity from the smeared distribution of quantities computed directly from the detector response. In machine learning, the same technique is known as quantification learning. In this talk, the concept and mathematical basis of unfolding is presented with examples solidly rooted both in the worlds of physics and machine learning.

Speaker: Lukas Liland (TU Dortmund)

8 Anomaly detection, BDTs and you: How the statistical properties of Boosted Decision Trees interact with physics

In particle physics, papers on weakly supervised searches for new physics have started accumulating numbers. This is the paradigm of trying to learn the differences between a data-driven background estimate and real data, in order to isolate some small contribution of anomalous data points - the signal - in the real data. The technique has been shown to work in simulation, and has been applied in several studies by experiments at the Large Hadron Collider. In recent years, boosted decision trees, or BDTs, have been shown to out-perform neural networks for binary classification in weakly supervised search contexts. They have several advantages; for example, training is faster, and they are more robust to noisy or uninformative features in the data.

However, BDTs also have some quirks. They work by partitioning the feature space of the data into varyingly sized, axis-aligned hyperboxes. This discretization can be considered unphysical, at least in the sense that physics is considered smooth down to the quantum level. Additionally, this partitioning is dependent on the complexity of the BDT as well as the amount and density of data. Combined, this has consequences for which data is actually considered "anomalous", in a way that may not be visible in high-level quantities like receiver operating curves and significance improvement characteristics.

In this talk, I dive down the rabbit hole of BDT anomaly detection as I've encountered it in my work studying high-energy physics data. I will visualize the choices a BDT makes in different conditions, which may or may not be in line with physics goals, and try to understand those choices in terms of statistical properties.

Speaker: Sten Åstrand (Lund University)

9 MultiCWoLa

In many classification problems, reliable instance-level labels are unavailable. However, it is often possible to construct weakly enriched unlabeled samples: datasets selected by different cuts, sources, populations, or experimental conditions that change latent class proportions without revealing them. Classification without Labels (CWoLa) shows that, in the binary case ($K=2$), a classifier trained to distinguish two impure mixtures with different class proportions can recover an optimal class discriminator without knowing the mixture proportions. We extend this principle to multiclass learning from several unlabeled mixtures ($K>2$), where the learner observes only mixture identity and neither latent class labels nor class-prior matrices. We prove that, for a multiclass mixture model, the Bayes-optimal mixture classifier $g^\star$ maps data points into a $(K-1)$-simplex embedded in mixture-posterior space. The $K$ vertices of this simplex are induced by the latent classes through the unknown mixing matrix. Leveraging this geometry, we propose prior-free procedures that train a standard classifier to distinguish mixture identities and then extract latent class structure using either post-hoc simplex fitting or a bottleneck architecture. Experiments on MNIST, CIFAR-10, and Galaxy10 DECaLS show that mixture identity alone can recover latent classes and their fractions in the mixture. By narrowing the gap between weakly supervised and fully supervised performance, we provide a mathematically grounded, scalable tool for multiclass discovery in label-scarce domains.

Speaker: Johann Ioannou-Nikolaides (NBI)

12:25 Lunch

10 Keynote 2: Cecilia Clementi

14:20 Coffee

Plenary: Molecules, matrials, structure

11 From AI-Assisted Characterization to Autonomous Discovery of Novel Ferroelectric Materials

Recent advances in large language models (LLMs), foundation models, and AI agents are transforming materials science from AI-assisted characterization toward autonomous scientific discovery. While machine learning has achieved remarkable success in materials property prediction, discovering entirely new functional materials remains challenging because of the enormous chemical design space and the fragmented nature of scientific knowledge distributed across literature and materials databases.

In this work, we present an AI-assisted framework for the autonomous discovery of novel ferroelectric materials by integrating automated literature mining, large language models, structured materials databases, and physics-guided machine learning. We have established a comprehensive ferroelectric materials database containing hundreds of unique compounds and thousands of extracted materials-property records, including crystal structures, spontaneous polarization, Curie temperatures, coercive fields, dielectric constants, piezoelectric coefficients, and other key descriptors. Large language models are employed to automatically extract and organize scientific knowledge from the literature, while materials databases such as the Materials Project provide complementary structural and computational information.

Building upon this knowledge base, we are developing a physics-guided AI discovery pipeline that combines LLM-based knowledge extraction with data-driven screening to identify promising unexplored ferroelectric candidates. Rather than relying solely on black-box prediction, the framework incorporates physical constraints—including crystallographic symmetry, thermodynamic stability, and established ferroelectric mechanisms—to improve both interpretability and prediction reliability.

This work represents a step toward autonomous materials discovery, where AI assists not only in analyzing experimental data but also in generating hypotheses and identifying new candidate materials for computational and experimental validation. More broadly, the framework provides a scalable approach for accelerating the discovery of functional quantum materials by integrating scientific literature, materials databases, machine learning, and physics-based reasoning.

Speaker: Prof. YAPING QI (Tohoku University)

12 Exploring hidden structure in water using polarizable atom interaction neural networks

Water is a key-ingredient for life as we know it and plays a central role in maintaining Earth’s climate within the limits that life can tolerate. Water is all around us, yet its properties are radically different from essentially all other substances. The density maximum at 4 degrees, that makes ice cubes float in our drinking glasses, is the most famous example among water’s anomalous properties. Thermodynamic response functions including isothermal compressibility and heat capacity at constant pressure are related to fluctuations in volume and entropy, respectively. These fluctuations increase upon cooling of ambient water and diverge in the supercooled regime.[1] An accumulating number of studies, both simulations[2] and experiments[3], show that the origin of these non-thermal fluctuations is a transition between two distinct forms of liquid water (a high- and a low-density liquid) and the existence of a liquid-liquid critical point.[4] The consequences of this bimodality are not limited to the supercooled regime but also affect the properties of ambient water in our drinking glasses and even the water in our cells. In my presentation I will explain how we use polarizable atom interaction neural networks to explore these fascinating phenomena and search for the structural motifs responsible for water’s unique properties. Is there hidden structure in ambient water?

References
1. P. Gallo et al., ”Water: A Tale of Two Liquids”, Chem. Rev. 2016, 116, 13, 7463-7500.
2. J. C. Palmer et al., ”Advances in Computational Studies of the Liquid–Liquid Transition in Water and Water-Like Models”, Chem. Rev. 2018, 118, 18, 9129-9151.
3. K. H. Kim et al., ”Experimental observation of the liquid-liquid transition in bulk supercooled water under pressure”, Science, 2020, 370, 978-982.
4. S. You et al., ”Experimental evidence of a liquid-liquid critical point in supercooled water”, Science, 2026, 391, 1387-1391.

Speaker: Dr Fredrik Grote (Department of physics, Stockholm University)

13 Predicting Magnetic Interactions from Time-Series Magnetization Data Using Explainable Machine Learning

Machine learning methods have proven useful for addressing a wide range of tasks in science and technology. Recent applications in materials science include materials discovery, predicting material properties, such as the Curie temperature, constructing phase diagrams, accelerating first-principles calculations, and more. One of the challenges in magnetism theory is accurately determining the magnetic interactions of a system to describe experimental observations and obtain a reliable description of the behavior of magnetic materials. To achieve this, density functional theory (DFT) is typically the method of choice. The resulting magnetic interactions are then used in further simulations employing Monte Carlo, atomistic spin dynamics, or micromagnetic methods. However, the calculated interactions may differ significantly depending on the DFT approach used, making benchmarking against experimental results more difficult.

Recently, researchers have begun extracting exchange interactions directly from experimental data using magnetic images instead of calculating them [1]. Using this approach, the Dzyaloshinskii-Moriya interaction has been extracted with good accuracy and validated through experimental measurements, and several magnetic interactions have been determined simultaneously. Due to the scarcity of experimental data and the difficulty of generating sufficiently large and systematically varied experimental datasets, simulated data have been used to train machine learning models, which are then validated using experimental images.

However, in addition to images, other information is available, including time-series data from so-called pump-probe experiments [2]. In our approach, we use time-series data rather than images to predict the magnetic interactions of various materials and multilayers [3]. Importantly, we address the black-box problem by applying machine-learning explainability methods to assess how the model makes decisions and, consequently, how likely it is to perform well on unseen data. We also examine the information used by the model and evaluate whether it can be understood from a physics perspective.

[1] M. Kawaguchi, et al., Npj Comput. Mater. 7, 20 (2021).
[2] M. Pankratova, et al. Scientific Reports 14, 8138 (2024).
[3] M. Pankratova, et al., Journal of Physics: Condensed Matter. 36 (2024)

Speaker: Maryna Pankratova (Uppsala University)

14 Predicting Superconductivity using Machine Learning

Room temperature Superconductors are one of the most exciting materials in condensed matter physics, but the hunt for such a material still remains ongoing. The physics of known high temperature superconductors are not well understood and the mechanisms vary and depends on chemical composition and crystal structure. To circumvent the problem of understanding the physics, we instead make use of ML methods to predict whether a material is superconducting and what the critical temperature is. Using the SuperCon [1] dataset which contains 12.000 chemical formulas and their corresponding critical temperature we make use of BDT's to predict properties of materials in the Materials Project [2]. We also attempt to use the crystal structure to improve our predictions using Graph Neural Networks trained on the 3DSC dataset [3].

[1] Materials Database Group. MDR SuperCon Datasheet Ver.220808. https://doi.org/10.48505/nims.3837

[2] A. Jain, S.P. Ong, G. Hautier, W. Chen, W.D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, K.A. Persson (*=equal contributions)
The Materials Project: A materials genome approach to accelerating materials innovation
APL Materials, 2013, 1(1), 011002.
doi:10.1063/1.4812323

[3] Sommer, T., Willa, R., Schmalian, J. et al. 3DSC - a dataset of superconductors including crystal structures. Sci Data 10, 816 (2023). https://doi.org/10.1038/s41597-023-02721-y

Speakers: Ms Amalie Falkenberg Davidsen, Ms Frida Birkedal Nielsen, Mr Hugo Schreckenberg (University of Copenhagen), Rasmus Madsen (University of Copenhagen), Silas Billeschou Schack (University Of Copenhagen)

15 High-Throughput Discovery of Magnetic Materials: By Machine Learning and its combination with First-Principles Methods

High-Throughput Discovery of Magnetic 2D Materials:
From Physics-Informed Machine Learning to Correlated Electron Systems

The discovery of magnetic two-dimensional (2D) materials has created new opportunities for spintronics, quantum technologies, and low-dimensional correlated electron physics. However, the enormous chemical and structural search space of magnetic monolayers makes conventional first-principles exploration computationally demanding. In this work, a high-throughput framework combining density functional theory (DFT), machine learning (ML), and physics-informed descriptors is presented for accelerated discovery and characterization of magnetic materials.

Transition-metal-based TMXY monolayers are investigated using supervised learning approaches including neural networks, random forest, and gradient boosting models. Key physical properties such as Hubbard (U), lattice parameters, magnetic exchange interactions, magnetic phases, and transition temperatures are predicted. Particular attention is given to descriptor engineering, where chemically and structurally motivated atomic fingerprints are employed to encode local environments, electronic occupation, and superexchange pathways governing magnetism in low-dimensional systems.

Beyond conventional high-throughput screening, recent developments toward integrating experimental and computational datasets within probabilistic data-assimilation frameworks for magnetic materials are discussed. Such approaches enable uncertainty-aware prediction and improved generalization for sparse materials datasets, particularly in rare-earth permanent magnets and correlated magnetic systems. In addition, interpretable machine-learning techniques, including feature-importance and SHAP-based analyses, are utilized to identify physically relevant descriptors controlling magnetic behavior.

Finally, the limitations of standard DFT and purely data-driven approaches in describing finite-temperature magnetism and electronic correlations are discussed, motivating future integration with many-body methods such as DFT+DMFT. The presented framework demonstrates how machine learning, first-principles calculations, experimental feedback, and correlated-electron physics may be combined for next-generation magnetic materials discovery.

Speaker: Atefe Nayamadi Mahmoodabadi (Uppsala)

16 Gaussian sum smoothing for continuously monitored quantum systems

The wave function encodes our knowledge of a quantum system prior to a measurement. By incorporating posterior information, one can retrospectively improve the estimate of unknown outcomes of earlier measurements. For a continuously monitored system, this inference problem is described by a Hidden Markov model whose forward and backward dynamics are governed by stochastic master equations. As quantum technologies continue to scale toward increasingly macroscopic systems, solving these equations becomes computationally prohibitive due to the curse of dimensionality associated with large state spaces.

In this contribution, we present a Gaussian-sum approach that circumvents this limitation. The generally non-Gaussian Wigner function is approximated by a quasi-mixture of Gaussian functions, each evolving according to linear equations of motion. For homodyne and heterodyne detection, the evolution of each Gaussian function can be solved independently and is therefore ideally suited for parallel computation. This enables efficient smoothing of higher-order moments in macroscopic bosonic systems.

Speaker: Julien Pinske (University of Rostock)

17 Generating pair distance distributions from SAXS data using machine learning

Small-angle X-ray scattering (SAXS) is widely used to determine the structure of nanoparticles and biomolecules in solution. A central challenge is reconstructing the pair distance distribution function, $p(r)$, from the measured scattering intensity, $I(q)$. The $p(r)$ function describes the distribution of distances between scattering centers within a particle and provides valuable information about its size and overall shape. Traditionally, $p(r)$ is obtained using indirect Fourier transform methods, which rely on iterative optimization and regularization.

In this project, we investigate whether machine learning can learn the relationship between $I(q)$ and $p(r)$ directly. Using a large dataset of simulated SAXS data generated with Shape2SAS, we trained and compared a multilayer perceptron (MLP) and a one-dimensional convolutional neural network (1D CNN) to predict $p(r)$ from scattering curves. We also examined how shape complexity and model architecture affect reconstruction accuracy.

Both models successfully recovered physically meaningful pair distance distributions, while the 1D CNN consistently produced the most accurate predictions, particularly for more complex particle shapes. Our results demonstrate that machine learning can rapidly recover structural information from SAXS data and highlight its potential as a complementary tool for SAXS analysis.

Speakers: Mr Ali Hussein (Niels Bohr Institute University of Copenhagen), Niklas Sørensen (Københavns Universitet)

Poster session and reception

Thursday 20 August

09:00 Coffee

18 Keynote 3: Shah Rukh Qasim

10:05 Coffee

Plenary: Experimental physics and detectors

19 Building foundation models with low-level collision data in particle physics

The integration of foundation models in particle physics is gaining pace rapidly and has expanded the search for new physics. This talk presents foundation models trained on low-level data from the first fully simulated dataset using Open Data Detector (ColliderML), to distinguish between Standard Model and Beyond Standard Model processes. We compare new physics discovery using only low level tracking information with the inclusion of more sources, such as different scales (tracks, particle flow objects) and different detector regions (calorimetry). The effect of these additions in both the supervised and unsupervised case is presented. The talk will further discuss the performance of these approaches in anomaly detection, highlighting the potential of pre-trained foundation models on low-level data for various downstream tasks.

Speaker: Shreya Saha (Adelaide University)

20 Detector simulation for neural network position reconstruction in the i-RASE system

The MeV gamma-ray domain suffers from poor sensitivity, motivating advanced detector and readout technologies. The i-RASE (Intelligent RAdiation SEnsor readout systems) project addresses this by combining a custom low-noise front end with FPGA-implemented, physics-informed neural networks for real-time event reconstruction in 3D CdZnTe (CZT) Drift-Strip Detectors (DSD). At DTU Space, work focuses on 3D CZT DSD design specifications, development of a physics-based detector model, and experimental validation. The framework integrates COMSOL signal simulations and Geant4 charge-cloud generation to optimize detector geometry and generate synthetic training data. Laboratory tests show that neural-network reconstruction achieves sub-millimetre FWHM spatial resolution at 662 keV, matching or exceeding conventional methods, and reduces dead zones by improving boundary position estimates.

Speakers: Michał Kossakowski (DTU Space), Rikke Stougaard Klausen (DTU Space)

21 Differentiable End-to-End Optimization of In-Ice Radio Neutrino Detectors

The design choices made for particle physics detectors can have long-lasting impacts on the scientific outcome of a given experiment. Optimizing detector layouts is therefore essential in the planning stages of an experiment. However, conventional brute-force Monte Carlo–based studies often become computationally prohibitive when exploring large design parameter spaces. Differentiable programming offers a solution by enabling calculation of gradients of scientific performance metrics with respect to detector design parameters, allowing efficient gradient-based optimization.
In this contribution, we present a fully differentiable end-to-end optimization pipeline for in-ice radio neutrino detectors targeting ultra-high-energy neutrino observations. The framework combines differentiable PyTorch implementations of radio signal generation, propagation, detection, and reconstruction using machine-learning-based surrogate models and uncertainty estimation through the Fisher information. This enables direct optimization of detector performance metrics with respect to antenna positions and orientations. We show proof-of-concept studies demonstrating end-to-end detector optimizations aimed at improving reconstruction precision for the IceCube-Gen2 radio array.

Speaker: Martin Ravn

22 Photonic imaging, sensing, and detection assisted by AI

Light carries rich physical information, but each detected photon can be costly in optical power, sample dose, or acquisition time. Modern photonic sensing, quantum technologies, and nanoscale imaging therefore require not only high sensitivity, but also maximal information efficiency. This talk reviews recent results from our group on AI-assisted photonic sensing and quantum detection, with the common goal of making each photon as informative as possible while reducing explicit calibration to a minimum.

I will first introduce device-agnostic learning for super-resolution emitter imaging. The aim is to use a single AI model across different optical devices and imaging conditions, avoiding the usual dependence on detailed imaging device calibration. We will discuss device-agnostic super-resolution of point-like sources in astronomy, fluorescence microscopy, stochastic single-molecule localization microscopy, and quantum-dot imaging [1]. Particular attention will be given to calibration-free emitter localization and reconstruction, including the prospects of combining device-agnostic imaging with photon-statistics detection for high-density and quantum-emitter applications [2].

The second part will focus on quantum entanglement detection and characterization. We will show how deep neural networks can quantify entanglement from incomplete random measurement projections, often outperforming standard tomography-based approaches. This framework naturally motivates device-agnostic entanglement quantification and the characterization of entangled emitters via selected quantum features, rather than through full state reconstruction [3]. Recent results on the scaling of measurement costs for entanglement quantification will also be briefly discussed.

Finally, I will discuss direct measurement of quantum features of light. Using random scattering in a short segment of few-mode fiber, multiplexed single-photon detection, and deep-learning-based decoding, we demonstrate an all-fiber polarization sensor operating at the single-photon level [4]. Its resource efficiency, performance, and relation to extreme learning machines will be reviewed, together with preliminary extensions toward multi-qubit direct detection and broader quantum-feature extraction.

References:

[1] D. Vašinka, et al., From stars to molecules: AI guided device-agnostic super-resolution imaging, arXiv:2502.18637, accepted in Nat. Commun. (2026); A. Dostálová, et al., Calibration-free single-frame super-resolution fluorescence microscopy, arXiv:2505.13293 (2025); D. Vašinka, et al., Universal super-resolution framework for imaging of quantum dots, arXiv:2510.06076 (2025).
[2] I. Straka, et al., Quantum non-Gaussian multiphoton light, npj Quant. Inform. 4, 4 (2018); J. Hloušek, et al., Accurate detection of arbitrary photon statistics, Phys. Rev. Lett. 123, 153604 (2019); J. Hloušek, et al., High-resolution coincidence counting system for large-scale photonics applications, Phys. Rev. Applied 21, 024023 (2024).
[3] D. Koutný, et al., Deep learning of quantum entanglement from incomplete measurements, Sci. Adv. 9, eadd7131 (2023).
[4] M. Bielak, et al., All-fiber microsensor of polarization at single-photon level aided by deep-learning, Laser Photonics Rev. 20, e01775 (2026).

Speaker: Miroslav Ježek (Palacký University Olomouc)

23 Performance of the Boosted Decision Tree for $\pi^0$ Reconstruction in KLOE Data

We present a Gradient-Boosted Decision Tree (BDT) for $\pi^0$ reconstruction in the $e^+e^- \to \pi^+\pi^-\pi^0\gamma$ process at the KLOE experiment. Trained on Monte Carlo events using ten kinematic variables and a series of preselection criteria (including $\chi^2 < 100$), the model is optimized via Bayesian hyperparameter search. On independent validation samples, the BDT achieves an AUC of 0.9989, a pair-level signal efficiency of 98\%, and a background rejection of 99\%. On the test set, optimization over aggregation strategies selects a threshold of 0.35 with the `mean' strategy, yielding an event-level signal efficiency (recall) of 99.7\%, a purity (precision) of 99.2\%, an accuracy of 98.9\%, and an F1 score of 0.99. Notably, the correct diphoton pair is identified in $>99.9\%$ of signal events, effectively removing combinatorial ambiguity. The tagger exhibits negligible overfitting, with a training--validation AUC gap of $5 \times 10^{-4}$. This MC-validated tagger is currently being deployed on real KLOE data. While the exact performance on data is subject to ongoing validation, these preliminary results promise substantial improvements over conventional cut-based selections.

Speaker: Dr Bo Cao (Department of Physics and Astronomy; Nuclear Physics, Uppsala University)

24 Enhancing Detector Alignment with Machine Learning

Precise detector alignment is essential for reliable charged-particle tracking in high-energy physics experiments. Standard track-based alignment methods typically depend on iterative global fits which can become computationally demanding for large and highly segmented detectors.

We investigate a machine-learning approach in which detector sensors are represented as nodes in a graph, while reconstructed particle tracks define graph instances. The model is trained to use track-sensor correlations to identify detector elements that are likely to be misaligned, providing an alternative to conventional hierarchical alignment procedures.

In a proof-of-concept study, the method identifies single misaligned layers with an accuracy above 99% and shows encouraging behaviour in multi-layer misalignment scenarios. The approach also allows alignment quality to be scored at the sensor level, which could make it useful for detector performance monitoring and, potentially, near-real-time alignment diagnostics.

Speaker: Stefan Cretu (Lund University)

25 Challenges of using machine learning in rare heavy-ion measurements

Searching for rare B meson decays in heavy-ion collisions requires separating a small signal from a large and highly variable background. Machine-learning methods could improve this separation, but their training would rely largely on simulated signal and background samples that do not fully reproduce the real data. Differences in detector response, event multiplicity, and background composition can therefore lead to domain shift and poorly controlled selection biases.

Using an analysis based on conventional selection methods as a case study, I discuss where machine learning could improve sensitivity, why its application is particularly challenging in high-density collision data, and what validation would be needed before such methods could be used in a precision measurement.

Speaker: Ida Storehaug

26 Machine Learning Extraction of Flow Harmonics and Jet Amplitudes from Two-Particle Correlations in Pb-Pb Collisions at √(〖s_NN〗_ ) = 2.76 TeV

We present a machine learning approach for the simultaneous extraction of anisotropic flow harmonics (v₁–v₆) and jet amplitudes from two-particle angular correlation functions in heavy-ion collisions. A multi-layer perceptron (MLP) neural network is trained on HIJING-like toy model simulations incorporating realistic physics ingredients, including collective flow parametrizations from ALICE data, jet fragmentation functions, thermal background spectra, and resonance decay contributions. The network takes as input 2D correlation histograms in (Δφ, Δη) space along with event-level observables (centrality, trigger and associated particle transverse momenta) and outputs the flow harmonic coefficients and near-side/away-side jet yields. The model is validated against ALICE HEPData measurements of trigger-associated correlations in 0–10% central Pb-Pb collisions at √(〖s_NN〗_ ) = 2.76 TeV with trigger pt in the range 8–16 GeV/c and associated pt between 1–6 GeV/c. We demonstrate that the MLP successfully decomposes the raw correlation function into its flow-modulated background and jet contributions, providing a model-independent alternative to traditional Fourier decomposition and template fitting techniques. The inclusion of non-linear hydrodynamic mode coupling for higher-order harmonics (v₄, v₅, v₆) is shown to be essential for accurate description of the data. This approach opens new possibilities for automated, high-precision extraction of jet quenching observables in the presence of strong collective flow backgrounds.

Speaker: olja dordic (university of banjaluka)

12:10 Lunch

Train, conference dinner, excursion

Friday 21 August

09:00 Coffee

27 Keynote 4: Gabriel Facini

10:05 Coffee

Plenary: Astrophysics and cosmology

28 Simulation-Based Inference of High-Redshift Galaxy Environments with JWST

The astrophysics of the first galaxies is undergoing a data revolution, driven by the infrared capabilities of JWST and the coming arrival of Roman, SKA, and other facilities. Yet inferring the physical processes governing the formation and evolution of the earliest galaxies remains difficult: these objects are distant and intrinsically faint, and several well-established scaling relations appear to break down at the highest redshifts, demanding revisions to our current models. This motivates a different approach to extracting knowledge from data — simulation-based inference (SBI), which forward-models the full observational process and yields posteriors without requiring an explicit, tractable likelihood.
We apply this approach to two problems in high-z galaxy formation. First, we build an SBI framework that forward-models entire JWST pointings around bright z > 10 galaxies, using a neural ratio estimator to produce fast, full posteriors on population parameters — luminosity boost versus increased stochasticity in star formation — that standard tools like the UV luminosity function cannot disentangle. Because the model is physically motivated, posteriors from multiple independent pointings combine naturally, and other observables such as the UVLF can be folded directly into the likelihood. Applied to real JWST data, this yields joint constraints favoring a picture of increased stochasticity in early star formation.
Second, we extend the same SBI framework to infer the size and shape of ionized bubbles around z > 6 Lyman-α emitters directly from JWST spectra, replacing likelihood-map approaches with full posteriors and properly modeled non-Gaussian covariances between galaxies.
Together, these applications demonstrate SBI as a general, scalable route to extracting astrophysical information from JWST's high-redshift galaxy environments.

Speaker: Ivan Nikolic (Dawn Center, University of Copenhagen)

29 Genesis-Metallicity: Non-Parametric Gas-Phase Metallicity Estimation for Galaxies Across Cosmic Time

The chemical enrichment state of galaxies is a result of a complex interplay between several mechanisms that (i) accrete pristine gas onto galaxies, (ii) cool down, fragment, and convert this pristine gas into stars, (iii) produce the heavy elements in the cores of massive stars, and (iv) release the newly formed heavy elements into the interstellar medium through stellar winds and supernovae. As such, scaling relations between gas-phase metallicity and other galaxy properties such as stellar mass and star-formation rate provide some of the most readily testable predictions of theoretical models and hydrodynamical simulations that can be directly calibrated against observational data. This has placed the measurements of gas-phase metallicity at the forefront of galaxy formation and evolution studies. The full array of faint emission lines required for precise gas-phase metallicity measurement is only available for a prohibitively small subsample (<1%) of spectroscopically targeted galaxies. Instead, strong emission-line ratios are used as proxies for gas-phase metallicity in large samples. For over three decades, the state-of-the-art has consisted of fitting polynomials to two-dimensional projections of the intrinsically multi-dimensional relation between gas-phase metallicity and strong emission-line ratios; this calibration is commonly carried out on samples where faint emission lines, and hence precise gas-phase metallicities, are available. This approach comes with several caveats, including (i) losing information as a result of projecting the complex parameter space onto two-dimensional planes, and (ii) saturation points where the gas-phase metallicity has a ~0.5-1.0 dex spread for little variation in the emission-line ratio. In this talk, I will present a novel approach where the full multi-dimensional relation between gas-phase metallicity and strong emission-line ratios is captured non-parametrically through kernel density estimation. This kernel density estimation is calibrated on the largest compilation of observational data to date, spanning distant galaxies in the early-universe to present-day galaxies. I will show that this approach achieves competitive accuracy in comparison to traditional techniques. Genesis-Metallicity is publicly available at: https://github.com/langeroodi/genesis_metallicity

Speaker: Dr Danial Langeroodi (University of Cambridge)

30 Multimodal Generative Modelling of the Galaxy Population with pop-cosmos

Panchromatic surveys such as COSMOS have expanded our understanding of how galaxies evolve through cosmic time immensely. Surveys such as the Vera C. Rubin Observatory's LSST are promising an unprecedented view of this evolutionary paradigm provided the massive data challenge they pose is answered. We have been developing pop-cosmos, a comprehensive galaxy population model housing a 16-parameter SPS model parametrized by a flexible diffusion model. The 16D distribution over galaxy properties including their redshift is forward-calibrated on 26-band photometry from COSMOS with a deep infrared selection. In a recent paper we used a galaxy catalogue drawn from the trained pop-cosmos model to investigate the stellar mass assembly and star formation histories of galaxies up to z=4. I will briefly summarize key results such as the cosmic star formation rate density inferred from our model, and will present a look into the quenching mechanisms of galaxy populations. Our investigation finds a shift of the cosmic dawn towards earlier lookback times, and uncovers correlations between star formation, AGN activity and quenching difficult to capture without a population-level analysis. In this talk I will mainly focus on our work expanding the forward process of pop-cosmos to jointly train on the morphology of galaxies and their photometry from profile fitting. The morphology of galaxies encodes important information about their evolutionary stage, and I will talk about how the inclusion of this dimension in the forward modelling impacts the population model. I will describe how we connect the morphology information using the compressed latent representation learned by an encoder-decoder network, like a convolutional variational autoencoder, effectively keeping the computational cost increase due to this new mode to a minimum. To conclude, I will present how this approach unlocks a new avenue for population-level causal inference in astrophysics, using the causal relation between quenching and morphological transformation as the primary example.

Speaker: Dr Sinan Deger (University of Cambridge)

31 Searching for Technosignatures with Machine Learning: The Breakthrough Listen LOFTS Dual-Station Radio Survey

Under the assumption that other advanced extraterrestrial civilisations exist - or have existed - elsewhere in our and other galaxies, the Breakthrough Listen Initiative has for the past decade taken the lead in the search for both direct and indirect technosignatures using a diversity of radio telescope facilities. Advances in instrumentation, computational processing power and most especially in ML have provided a means for researchers to both efficiently and rigorously search for anomalous signals against a continuing deluge of complex and very noisy observational radio data. One such Breakthrough Listen project, LOFTS (LOw Frequency pulsar, FRB, and Technosignature Survey) has been using two LOFAR (LOw Frequency ARray) outstations in Ireland and Sweden to simultaneously observe several galactic regions encompassing over a million nearby main sequence stars (mean distance of 1215 pc) at VHF radio frequencies. Being so geographically far apart, any locally bright sources of radio frequency interference can be identified and removed when the data from both stations are combined, with any remaining features common to both likely to be non-terrestrial in origin. I describe how we are developing a machine-learning framework that can expeditiously distinguish true dual-station coincidences from site-local RFI for broadened signal morphologies consistent from distant Doppler varying and locally stellar space weather distorted radio beacon signals. I place this work against the wider use of ML in the Breakthrough Listen Initiative to date, and the impact this methodology continues to make in resolving arguably one of the most profound questions for our species, whether we on our home planet are the sole sentient witnesses to our Universe.

Speaker: Prof. Aaron Golden (University of Galway)

32 Stellar Stream Identification with Hyperparameter Transfer GNNs

Stellar streams, the remnants of tidally disrupted dwarf galaxies and globular clusters, are powerful probes of galaxy assembly and dark matter. While many streams have been discovered in the Milky Way, comparisons with cosmological simulations suggest that both the census of streams and the identification of their constituent stars remain incomplete due to overwhelming stellar backgrounds, diverse stream morphologies, and partial or uncertain observational data. We present a machine-learning framework for stellar stream identification based on graph neural networks (GNNs) tuned for hyperparameter transfer, a strategy for parametrizing NNs so that near-optimal hyperparameters identified in small, cheap-to-optimize models, are the same for their larger counterparts. Unlike traditional cut-based analyses or those that rely on explicit models of the Galactic potential, GNNs learn local structures by message passing across stars embedded as graph nodes. This relational inductive bias makes them well suited to the detection of faint tidal features, diffuse streams, and irregular substructures. We apply this framework to stellar populations in the Dark Energy Survey, with the goal of improving membership identification for known streams and establishing a scalable foundation for future stream discovery efforts with upcoming surveys such as LSST and Roman in the Milky Way and nearby galaxies.

Speaker: Adriana Dropulic (Niels Bohr Institute)

33 What Shapes the UV β Slope of Early Galaxies? An ML Approach to Galaxy Evolution in the Epoch of Reionization

The UV continuum slope (β) of high-redshift galaxies is an important observable for understanding the stellar populations, dust attenuation, and ionizing photon production of galaxies during the Epoch of Reionization. However, the relationship between UV β slopes and underlying galaxy properties is highly non-linear, involving the combined effects of star formation, metallicity, gas evolution, and feedback processes.

In this work, we develop a machine learning framework to identify the key physical properties that determine the UV β slope of reionization-era galaxies using the SIMBA-EoR cosmological simulation. We train supervised learning models to predict galaxy UV β slopes from a set of intrinsic galaxy properties, including stellar mass, star formation rate, metallicity, gas properties, and structural parameters. By comparing model performance and applying interpretable machine learning techniques, we quantify the relative importance of different galaxy properties and uncover the dominant features driving variations in UV spectral slopes.

Our approach allows us to explore complex non-linear relationships beyond traditional correlation analyses and provides a systematic method to connect observable galaxy properties with their underlying physical processes. The results will help improve the interpretation of high-redshift galaxy observations from facilities such as the James Webb Space Telescope and provide new insights into the physical origins of UV β slope diversity during cosmic reionization.

Speakers: Mr Davey Plugers (University of Copenhagen), Zhen Xiang (University of Edinburgh)

34 Transformers for heterogeneous time series of strongly lensed supernovae

When a supernova is located behind a massive galaxy, its light can be gravitationally lensed to form multiple images. Such a strongly lensed supernova is a rare and powerful probe that provides insights into distant supernova explosions, dark matter in galaxies, and even the expansion rate of the Universe. Currently, the lensed supernova field stands at a turning point, as we transition from a handful of known systems to several hundred with the advent of the Vera Rubin Observatory. However, this increase in data volume also introduces complex new challenges. In my talk, I will focus on the problem of performing fast inference of time delays from irregularly sampled, multi-band observations. I handle this heterogeneous data using a transformer with rotary positional embeddings, combined with simulation-based inference. I will share results from this work in progress, and invite discussions about handling irregular, multi-channel time series data across different physics domains.

Speaker: Nikki Arendse (University of Nova Gorica)

12:05 Lunch

Plenary: Climate, Earth, Environment

35 Deterministic Collapse and Stochastic Recovery in a Data-Driven Model of the Atlantic Meridional Overturning Circulation

During the last ice-age, temperatures in Greenland have frequently increased and decreased by 10$^{\circ}$C. Detailed studies with climate models suggest that this is caused by collapses and recoveries of the Atlantic Meridional Overturning Circulation (AMOC). The causes of these AMOC transitions are still debated. Here we describe the development of a neural-network based surrogate model of the AMOC. It is trained on 32,000 years of climate model integrations to build a set of stochastic differential equations that emulate the climate models' AMOC behavior. In particular it reproduces the spectra and the asymmetry in the times it takes for the AMOC to recover and collapse, which makes it more realistic than previously published sets of coupled differential equations to study past AMOC transitions. Monte Carlo simulations with this model show that collapses are deterministic, but recoveries are stochastically forced, in partial support of the leading hypotheses surrounding the AMOC transitions.

Speaker: Qi-fan Wu

36 Transferability and application of climate emulators to produce Antarctic ice sheet surface mass balance

Regional Climate Models (RCMs) bring significant added value compared to using global earth system models directly for modelling the Surface Mass Balance of the Antarctic Ice Sheet. However the high computational cost of RCMs limits their application across both the CMIP ensemble and in running different pathways. An alternative is the use of deep learning based climate emulators to expand the ensemble of SMB products available. In this study we first demonstrate the added-value of a multi-model ensemble of RCMs in estimating Antarctic SMB using the SUMUP database of in-situ and radar observations. We then use the output of one member (Harmonie-Climate, HCLIM) of the ensemble to produce downscaled SMB at high temporal and spatial resolution with a U-Net-based neural network. ERA5, a reanalysis nudged by observations, represents the actual observed weather and improves fidelity in the historical period, unlike the Earth System Models (ESMs) which have their own internal variability. The foundational training of the emulator is therefore based on ERA5-driven HCLIM model output, such that it learns the most realistic climate and SMB patterns. We then fine-tune the emulator on a variable number of years of HCLIM driven by the ESMs CESM2 and MPI-ESM. We evaluate the transferability of the emulator by estimating its performance as a function of the number of years used for fine-tuning. We also investigate the transferability to emissions pathways not available in the training dataset. Our emulation approach allows us to produce a wider ensemble of daily SMB values than currently available by RCM downscaling alone. We apply this technique to a range of several CMIP6 ESM members for the period 2015-2100 and quantify the representativity of using only one of the ensemble members to produce future projections of SMB.

Speaker: Clément Cherblanc (Niels Bohr Institute + Danish Meteorological Institute)

37 Interpretable Target-Aware Deep Learning: Linking Weather Patterns to Precipitation

Classical dimensionality reduction has served physics well. PCA is fast, transparent, and cleanly ranks structures by explained variance, which is why it remains the default first look for almost any high-dimensional dataset. But its strengths come from assumptions the underlying physics rarely respects, the modes must be linear and mutually orthogonal, forcing genuinely distinct physical states to be smeared across several components or merged into one. Unsupervised methods such as clustering or vanilla autoencoders relax these constraints, yet they introduce a different problem: they find whatever structure dominates the input, with no guarantee that it bears any relation to the quantity we actually care about. How do we then tease out non-linear, non-orthogonal, and non-dominant, but still relevant, structures in our data?

As a possible answer to this question, we present a Regression Mixture Model Convolutional Variational Autoencoder (RMM-CVAE). A CNN encoder compresses input fields into a low-dimensional latent space with a mixture-of-Gaussians prior, so that each mixture component corresponds to an interpretable regime. A regression head trained jointly with the autoencoder ties the latent space to a target variable, making the discovered regimes informative about impacts rather than merely about input similarity. Because the model is generative, each regime can be decoded back into physical space and inspected directly. Along the way, we discuss the practical lessons that transfer beyond our application: keeping mixture-prior VAEs stable during training and choosing the right output distribution for skewed, zero-heavy data.

As a demonstration, we apply the framework to a problem from climate physics: discovering large-scale atmospheric patterns from daily sea-level pressure fields that drive precipitation in Europe. The model recovers well-known circulation patterns, alongside regimes that linear methods cannot separate, and links each of them to its regional rainfall fingerprint. The same recipe should be applicable wherever interpretable latent states must be connected to physical outcomes.

Speaker: Mr Jonathan Ortved Melcher (University of Copenhagen, Niels Bohr Institute, Physics of Ice Climate and Earth)

38 From Micro to Macro-World: Device-Agnostic Single-Emitter Super-Resolution Imaging using Deep Learning

Any optical imaging system, from microscopes to telescopes, is subjected to diffraction owing to its finite aperture. For objects composed of many single emitters, such as molecules, quantum dots, or stars, diffraction blurring can significantly impair our ability to determine the number and positions of emitters present. Deep learning techniques showed great potential in overcoming this problem. However, these methods remain device-specific, requiring explicit knowledge of the imaging system and its parameters.
We propose a novel device-agnostic deep learning approach that surpasses these limitations. In this work, we demonstrate the generalization ability of the device-agnostic modeling on an unprecedented scale. From optical laboratory systems and dense-molecule microscopy to galaxy observations, our approach uses no information about the imaging system while achieving better accuracy than established methods.
Using a single device-agnostic model, without retraining or calibration, we reconstruct high-resolution images across numerous applications. Namely, we applied it to stacks of single-molecule localization microscopy data, revealing tubulin structures and nuclear pore complexes in great detail. Additionally, we verify the super-resolution ability by reconstructing fully unresolved binary stars from ground-based telescope data and comparing the results with those from a space observatory. Finally, we analyze generalization across numerous simulated imaging systems with widely varying parameters, such as signal-to-noise ratio, emitter density, and point-spread function profiles, including aberrations.

Speaker: Dominik Vašinka (Palacký University Olomouc)

39 Does extrinsic flux patterns support clustering by intrinsic oscillations in insect lidar?

What goes up doesn’t necessarily come down. No insects are born in the air but many ends their lives in bellies of predators. Over breeding sites this imbalance would be exaggerated and during invasions the flux could be reversed. Our group currently develops lidar tools for online in situ biodiversity assessment. We can detect hundreds of thousands of insects per day and differentiating hundreds of intrinsic oscillatory signal types by an unsupervised clustering criterion. Proving that clusters indeed represent different species and sexes is, however, complicated, laborious and costly. One concept in biology states that species must have different niches, functions and behave differently to coexist. We previously investigated how clusters differs in daily activity patterns. We now found multiple complimentary temporal asymmetries in our signals reflecting accent and decent rates during the day. We present imbalances and investigate to what extent these rates and imbalances are cluster specific and can further support our biodiversity criterion.

Speaker: Mikkel Brydegaard (Lund Univeristy)

40 Magnetic-field controlled perfect chirality in a on-chip coherent spin–photon interface

The ability to engineer chiral spin-photon interactions is essential for exploring quantum networks and non-reciprocal quantum optics. Here, we demonstrate chiral coupling of a negatively charged quantum dot spin, i.e., a four-level system, embedded in a standard W1 photonic-crystal waveguide. Despite the absence of global chirality, we observe spin-dependent directional emission of 0.77$\pm$0.02 and extract a chiral phase of 0.28$\pm$0.01$\pi$ in the Faraday magnetic fields. %By transitioning to the Voigt geometry, we characterize the intrinsic cyclicity of the inner and outer $\Lambda$-system transitions to be $\sim6$.
By tuning the magnetic-field orientation, we achieve dynamic control over the polarization of individual optical transitions, enabling perfect directional emission (0.99$\pm$0.01) with imperfect polarization. In a two-sided waveguide, this control allows a selected transition to decay exclusively into a single propagation direction. We further observe two tunable chiral cyclicities between two $\Lambda$-systems, with upper bound values reaching up to 134 (72.4–323). Finally, we show a proof-of-concept that this local chirality enables remote optical control of electron spin rotation via an optical Raman process mediated by waveguide transmission. This approach is expected to reduce the required driving power and suppress laser-induced spin decoherence. Building on the simultaneous realization of high cyclicity and directionality, we propose a single-click protocol with high loss tolerance for generating remote spin--spin entanglement. Our results establish the magnetic field as a powerful degree of freedom for engineering chirality in standard nanophotonic interfaces.

Speaker: Dr Shikai Liu (The Niels Bohr Institute, University of Copenhagen)

41 Outlier Detection in the BedMap Dataset

The BedMap compilation of Antarctic ice-thickness measurements contains 74.7 million data points collected over five decades (1966–2020). Measurement noise, unphysical jumps, and the ongoing ice flow over this period mean that much of the older data is unreliable, and such corrupted measurements propagate into the climate models built on top of the dataset. We present a semi-supervised graph neural network (GNN) approach for point-level outlier detection in this dataset. Pseudo-labels are generated from geometry alone: wherever two independent survey tracks cross within a small distance, the same spot has effectively been measured twice, and the two measurements are compared against each other and against the surrounding points within a local support radius, where depth changes and slope steepness between neighbours determine whether each measurement is physically consistent with its surroundings. Agreeing, consistent pairs are labelled inliers; measurements that contradict their partner and fall outside the locally allowed band are labelled outliers. All points are then connected in a k-nearest-neighbour graph whose edges carry relative distances and thickness gradients, and the GNN is trained on the pseudo-labels in two regional folds and evaluated cross-region. This evaluation on unseen seed labels achieves an AUC of approximately 0.86, and the model produces an outlier probability for all 74.7 million points, flagging 703k high-confidence outliers (p > 0.7). We discuss how geometry-derived pseudo-labels combined with spatial message passing may be used to clean large heterogeneous geophysical datasets, along with the limitations of the approach.

Speaker: Mr Janik Happel (NBI, KU)

42 Panel and Discussion

43 Outro

15:25 Coffee, Beer, Farewell