Neutrino Physics and Machine Learning (NPML 2025)

20 Oct 2025, 09:00 → 31 Oct 2025, 17:00 Asia/Tokyo

Koshiba Hall (University of Tokyo)

About the conference:

The 4th Neutrino Physics and Machine Learning  (NPML 2025) will take place in Japan. The NPML conference series are dedicated to identifying new opportunities, developing and sharing firm knowledge base, and building the future visions for impactful Artificial Intelligence and Machine Learning (AI/ML) research for neutrino physics. 

We look forward to your contributions to share the latest AI/ML research advancements at all levels of applications in neutrino physics, including experimental design optimization, detector operations and calibrations, physics simulations, data reconstruction, and physics inference.

We invite both individual speakers as well as representatives from a large collaboration in the neutrino community. Speakers from outside neutrino physics are also welcome to make contributions: your contributions will bring new insights and help us develop interdiciplinary research collaborations.

Key Information

• Registration Fee
  • Please pay the fee as indicated in the email you received here.
• Location: 
  • Main conference: Koshiba Hall, University of Tokyo
  • Satellite workshop: Seminar Room B, Kavli Institute of Physics and Mathematics for Universe (IPMU)
• Dates:
  • Main conference: October 27th to 31st
  • Satellite workshop: October 20th to 24th (tentative)
• Early registration (deadline June 30th 2025) : 
  • 30,000 JPY (regular registration)
  • 20,000 JPY (student registration)
• Standard registration (deadline July 31st 2025)
  • 40,000 JPY (regular registration)
  • 30,000 JPY (student registration)
• Financial support (deadline June 30th 2025)
  • Registration fee waivier
  • Accommodations 
  • Transportation (domestic travel only)

Contributing Talks/Posters:

Please indicate your interest in the registration form. You do not need to submit a formal title nor abstract at the time of registration - this is to motivate early registration as soon as possible. 

The submission deadline of a formal title and abstract is August 31st. When the official title and abstract are ready, please submit them from the Call for Abstracts page.

At NPML, we strongly encourage speakers of oral presentation to also consider a poster presentation which allows participants to interact more in depth with you and learn about your research. 

Satellite workshop:

We will hold a satellite data workshop in the week of October 20th at Kavli IPMU in Kashiwa-city, Chiba (30-40 minutes by Tsukuba Exp. rail). The workshop will host multiple events including the workshop toward the first Data Olympic for Neutrino Physics and Machine Learning (DO-NPML) and related working sessions to develop critical dataset for advancing AI/ML research and catalyzing interdisciplinary collaboration. 

The details of the satellite workshop can be found in a separate page (link). The goal of DO-NPML is to enable collaborative development of common solutions and benchmarks. Please join, develop, and train your AI/ML model and share with the community!

All participants are welcome to join the satellite workshop. Please indicate your interest in the registration form. Further details about DO-NPML will be shared soon.

Local Organizational Committee

• Patrick de Perio (IPMU/U-Tokyo)
• Masashi Yokoyama (U-Tokyo)
• Yasuhiro Nakajima (U-Tokyo)
• Kimihiro Okumura (ICRR/U-Tokyo)
• Yoshitaka Itow (ICRR/U-Tokyo)
• Benjamin Quilain (IN2P3/U-Tokyo)
• Aya Ishihara (Chiba-U)

International Organizational Committee 

• Patrick de Perio (IPMU/U-Tokyo)
• Kazuhiro Terao (SLAC)
• Saul Alonso Monsalve (ETH)
• Marta Babicz (U. Zurich)
• Jianming Bian (UC Irvine)
• Leigh Whitehead (Cambridge)
• Aobo Li (UCSD)
• Marco Del Tutto (Fermilab)
• Taritree Wongjirad (Tufts)
• Marco del Tutto (FNAL)
• Nick Prouse (Imperial)

Acknowledgements:

This conference is supported by Kavli IPMU, CD3, and Kakenhi International Leading Research (Grant #24K23938) funding. 

Updates:

• March 30th - the venue is decided.
• June 2nd - registration is open!
• June 23rd - LOC and IOC listed
• August 8th - Emails sent about financial support, support letter for a foreign visitor visa application, as well as confirmation of talks and posters are sent out.
• August 25th introduced several new indico pages.
  • Satellite workshop
  • Accommodation recommendation/discount at IPMU
  • Childcare support
• October 8th published the agenda 
  • Timetable 
  • What's unconference?

Monday 20 October

09:00 → 10:55 Satellite Unconference Kick-off

09:00 Welcome and logistics

Speaker: Patrick de Perio (Kavli IPMU)

09:10 Initiating the NPML data tracks

Speaker: Kazuhiro Terao (SLAC National Accelerator Laboratory)

09:20 Q/A

09:25 Example: Foundation Models as a multi-year data challenge

09:45 Q/A

09:50 Multi-modal event generator surrogate

Speaker: Callum Wilkinson (LBNL)

10:10 Q/A

10:15 Coffee

10:35 Unconference and Quad Charts

Speaker: Kazuhiro Terao (SLAC National Accelerator Laboratory)

10:50 Q/A

10:55 → 11:40 Unconference: How to organize datasets and challenges

10:55 Session Introduction

11:15 Q/A

11:20 Group formation

11:30 Workshop Photo

13:00 → 14:40 Unconference: How to organize datasets and challenges

13:00 Quad chart

Session Folder

13:45 Group presentation

14:20 Summary build

Summary Document

14:40 → 17:40 Unconference: Reconstruction

14:40 Introduction

14:55 Q/A

15:00 Coffee Break

15:30 Preparation

List of topics

16:00 Quad chart 1

Session Folder

16:45 Group presentation 1

17:20 Summary build 1

Summary Document

Tuesday 21 October

09:30 → 11:40 Unconference: Reconstruction

09:30 Preparation

09:40 Quad chart 2

Session Folder

10:25 Coffee break

10:45 Group presentation 2

11:20 Summary build 2

Summary Document

13:00 → 14:40 Unconference: Reconstruction

13:00 Quad chart 3

Session Folder

13:45 Group presentation 3

14:20 Summary build 3

Summary Document

14:40 → 17:40 Unconference: Detector physics modeling

14:40 Introduction

14:55 Q/A

15:00 Coffee break

15:30 Preparation

List of topics

16:00 Quad chart 1

Session Folder

16:45 Group presentation 1

17:20 Summary build 1

Summary Document

Wednesday 22 October

09:30 → 11:40 Unconference: Detector physics modeling

09:30 Preparation

09:40 Quad chart 2

Session Folder

10:25 Coffee break

10:45 Group presentation 2

11:20 Summary build 2

Summary Document

13:00 → 14:40 Unconference: Detector physics modeling

13:00 Quad chart 3

Session Folder

13:45 Group presentation 3

14:20 Summary build 3

Summary Document

14:40 → 17:40 Unconference: Analysis

14:40 Introduction

14:55 Q/A

15:00 Coffee break

15:30 Preparation

List of topics

16:00 Quad chart 1

Session Folder

16:45 Group presentation 1

17:20 Summary build 1

Summary Document

Thursday 23 October

09:30 → 11:40 Unconference: Analysis

09:30 Preparation

09:40 Quad chart 2

Session Folder

10:25 Coffee break

10:45 Group presentation 2

11:20 Summary build 2

Summary Document

13:00 → 14:40 Unconference: Analysis

13:00 Quad chart 3

Session Folder

13:45 Group presentation 3

14:20 Summary build 3

Summary Document

14:40 → 17:40 Unconference: data portal/challenge structure

14:40 Introduction

14:55 Q/A

15:00 Coffee break

15:30 Quad chart

16:25 Group presentation

17:00 Summary build

Friday 24 October

09:30 → 11:30 White paper writing

Monday 27 October

09:00 → 09:40 Conference Organization

09:00 Opening

Speaker: Patrick de Perio (Kavli IPMU)

09:15 NPML Conference Planning

Speaker: Kazuhiro Terao

09:30 Q/A

09:40 → 11:35 Experiments - SK/HK

09:40 Overview of reconstruction methods in Super and Hyper-Kamiokande

The detectors of the Kamiokande saga relies on a well-established detection method using Cherenkov effect in ultra-pure water. Consequently, the reconstruction algorithms, especially of the Super-Kamiokande detector, have been developed since several decades, allowing to reach unparallel results in neutrino sector, from the constraints on the solar neutrino upturn to the atmospheric neutrino constraint on mass ordering. This continuity in the detector has allowed an extremely rich development of the reconstruction algorithms, whether by refinement of existing methods or by introduction of new algorithms.

In this talk, we propose to review the reconstruction algorithms of the Super- and Hyper-Kamiokande experiments. In the context of the construction of Hyper-Kamiokande, an unprecedented development of Machine Learning-based algorithms has recently emerged. Though extremely powerful, these algorithms performances are often limited or over-tuned due to our unability to guide efficiently Machine Learning algorithms to identify key physics parameters. This talk propose to give an overview of the most traditional to Machine Learning based algorithms in order to overcome this challenge and maintain a fruitful dialog between the two approaches.

Speaker: Benjamin Quilain

10:05 Q/A

10:15 Graph Neural Networks for Hyper K Reconstruction

"The Hyper-Kamiokande Detector represents the next generation of neutrino observatories, following in the lineage of the Kamiokande and Super-Kamiokande experiments. With significantly enhanced sensitivity, Hyper-Kamiokande will support a diverse and ambitious physics program, including searches for proton decay, studies of solar neutrinos under non-standard scenarios, and the potential first observation of leptonic CP violation.
Designed to contain 260 kilotons of water and equipped with 20,000 photomultiplier tubes (PMTs), the scale and complexity of Hyper-Kamiokande necessitate the development of advanced event reconstruction algorithms. Existing techniques, originally developed for Super-Kamiokande, are beginning to show their limitations in this new experimental context.
In this presentation, we explore how next-generation approaches from the field of deep learning—specifically, Deep Neural Networks—can enhance reconstruction performance for Hyper-Kamiokande. Particular emphasis will be placed on the application of Graph Neural Networks (GNNs), presenting early promising results, and a performance comparison with the current reconstruction algorithms used in Super-Kamiokande adapted for Hyper-Kamiokande
"

Speaker: Erwan Le Blévec

10:30 Q/A

10:35 coffee

10:55 Improving Event Reconstruction in Hyper-Kamiokande with ResNet

A key goal of neutrino physics is to measure CP violation in neutrino oscillations, which may explain why the universe is dominated by matter over antimatter. The upcoming Hyper-Kamiokande experiment will be central to this effort. With a much larger detector volume and more photomultipliers than Super-K, Hyper-K will collect data at kilohertz rates. Traditional reconstruction tools are already at their computational limits, motivating new approaches. To this end, we will present our work on machine-learning-based reconstruction. We have developed an ensemble of ResNet models that improve particle momentum, position, and direction reconstruction, alongside a multi-class classifier to distinguish between key particle types. These networks achieve reconstruction and classification at speeds up to three to five orders of magnitude faster than existing techniques.

Speaker: Andrew Atta

11:10 Q/A

11:15 Vision Transformers for event reconstruction in water Cherenkov detector

Hyper-Kamiokande (Hyper-K) is a next-generation water Cherenkov neutrino experiment currently under construction, designed to address key questions in particle physics, including leptonic CP violation and proton decay. Convolutional neural networks (CNNs) have previously been applied to water Cherenkov detectors by treating PMTs as pixels, with charge and timing information serving as input features. In this work, we introduce an innovative application of the Vision Transformer (ViT) model. As an image-based architecture, ViT shows promising performance in event reconstruction, suggesting potential advantages over CNNs or graph neural networks (GNNs) in water Cherenkov detectors. In this talk, I will outline the ViT network architecture, discuss the motivation for exploring ViT and present initial results with SwinT, a variant of ViT, which indicate encouraging performance in water Cherenkov detector applications.

Speaker: Shuoyu Chen

11:30 Q/A

11:35 → 11:45 Group Photo

13:00 → 13:55 AI/ML in HEP

13:00 AI/ML for Accelerator Physics

TBD

Speaker: Daniel Ratner

13:40 Q/A

13:55 → 15:15 Experiments - SK/HK

13:55 The Water Cherenkov Test Experiment as a Demonstrator of Machine Learning Techniques for Neutrino Experiments

The Water Cherenkov Test Experiment (WCTE) is a 30-ton water Cherenkov detector that received 100-1200 MeV electrons, muons, charged pions and protons from the CERN East Area T9 beam, as well as observing secondary neutrons captured on Gadolinium and tagged photons through operation in a dedicated setup. With a suite of beamline detectors to characterise and tag the particles entering the tank, the WCTE is used to study the water Cherenkov detector response and physics interactions of particles typically produced in neutrino detectors. This provides a unique opportunity for novel technologies and techniques to be demonstrated on real data with known particle fluxes, towards reaching 1% level systematic uncertainties for GeV scale neutrino interactions. This talk will provide an overview of WCTE itself, its phyiscs goals and their potential to enhance future neutrino measurements. Central to this will be the machine learning based improvements to event reconstruction and detector calibration, with plans for demonstrating these new technologies on data collected over the past year, presenting opportunities for their first ever real-world validation with a water Cherenkov detector in a fully characterised testbeam.

Speaker: Nick Prouse

14:10 Q/A

14:15 WCTE Event Reconstruction with Graph Neural Networks

The Water Cherenkov Test Experiment is a powerful detector for studying water Cherenkov detectors, demonstrating new detection technologies testing reconstruction algorithms with controlled data. In particular, the relatively small scale of the detector allows for a fast and lightweight exploration of machine learning methods. We will present here the performances of Graph Neural Networks for event reconstruction - namely particle identification, vertex, energy and direction reconstruction - using the CAVERNS framework. Comparison with traditional reconstruction methods (FiTQun) is also shown.

Speaker: Mathieu Ferey

14:30 Q/A

14:35 Simulation, calibration and reconstruction in water Cherenkov detectors with machine learning techniques

Traditional Monte Carlo (MC) simulations rely on accurate modeling of both physics processes and detector responses. However, optimization with calibration data is often challenging,or even impractical, due to stochastic effects and the high dimensionality of correlated parameters.

Within the CIDeR-ML Collaboration, we are developing differentiable surrogate models that enable automated optimization to reduce data-MC discrepancies. In this contribution, we will present an application to the Water Cherenkov Test Experiment (WCTE), a test beam experiment at CERN that concluded operations in Spring 2025. We will also compare the performance of this approach with that of traditional calibration and reconstruction methods.

Speaker: Ka Ming Tsui

14:50 Q/A

14:55 Coffee

15:15 → 16:10 Experiments - T2K

15:15 Transformer-based Approaches to Particle Identification in the SuperFGD Detector.

Machine learning methods were explored for particle identification (PID) in the SuperFGD detector, which part of the recent upgrade to the T2K near detector. Baseline models were established to provide benchmarks, followed by a classical neural network where feature engineering was applied to improve separation. A transformer-based architecture was then developed, with data augmentation, class-weighting, and pre-trained models tested to address class imbalance and improve generalization. Ongoing studies aim to clarify the comparative performance of these approaches and their potential role in future PID pipelines.

Speaker: Kiseeva Viktoriia

15:30 Q/A

15:35 Dual use of Normalizing Flows for efficient neutrino-nucleus cross-section sampling and bayesian modelling of systematical uncertainties for the T2K near-detector fit

We present a dual application of conditional normalizing flows to (i) accelerate Monte Carlo sampling of exclusive neutrino–nucleus cross-sections and (ii) model uncertainty distributions within bayesian fits of systematics, illustrated on the T2K near-detector fit. Although aimed at different applications and uses, both efforts share a nearly identical implementation: an iterative training procedure that refines a conditional flow to match high dimensional conditional distributions. We describe the common framework, task-specific adaptations, and the resulting gains compared with Monte Carlo techniques currently used for these tasks.

Speaker: Mathias El Baz

16:00 Q/A

16:10 → 17:45 Experiments - JUNO

16:10 Cutting-Edge Machine-Learning Advancements at JUNO

The Jiangmen Underground Neutrino Observatory (JUNO) is the world's
largest 20-kiloton liquid scintillator (LS) detector in south China. It
will precisely measure the oscillation of reactor antineutrinos from two
commercial nuclear power plants 53km away, with the goal of determining
the neutrino mass ordering and measuring three oscillation parameters to
sub-percent precisions. This talk reviews the advancements in event
reconstruction of JUNO with an emphasis on cutting-edge machine-learning
developments.

Speaker: Benda Xu

16:35 Q/A

16:45 Reconstruction and Identification of Atmospheric Neutrino Events at JUNO Using Machine-Learning Methods

The Jiangmen Underground Neutrino Observatory (JUNO) is a next-generation multipurpose
liquid scintillator detector with a 20-kiloton target mass, located in southern China. One of its primary goals is to determine the neutrino mass ordering (NMO) with a significance of at least 3σ by precisely measuring the oscillation pattern of reactor antineutrinos over a 53 km baseline. Beyond reactor neutrinos, JUNO also provides a unique opportunity to study atmospheric neutrinos, which play an important role in enhancing the NMO sensitivity through matter effects. The construction of the detector has been completed, and JUNO has already started physics data taking. Detecting atmospheric neutrinos in a liquid scintillator detector poses several challenges, such as the reconstruction of complex event topologies and the separation of interaction channels. Advanced machine learning methods, in particular deep-learning–based reconstruction techniques, offer promising solutions to address these difficulties. This talk will present recent progress in using such methods to reconstruct the energy, direction, and vertex of atmospheric neutrino events, as well as their performance in particle identification from Monte Carlo studies, highlighting both the challenges and the advantages of these innovative approaches.

Speaker: Milo Charavet

17:10 Q/A

17:20 Simulation-Based Inference for Precision Neutrino Physics through Neural Monte Carlo Tuning

The Jiangmen Underground Neutrino Observatory (JUNO), which has recently begun operations, is an ambitious experiment designed for precision neutrino physics. To achieve its main objectives — the determination of neutrino mass ordering, and the sub-percent measurement of the θ_12, Δm_21^2 and Δm_31^2 neutrino oscillation parameters — the experiment demands highly accurate Monte Carlo (MC) simulations. These simulations must describe the complex response of the 20-kton liquid scintillator target within a 35.4 m diameter acrylic sphere, which is monitored by the 17,596 20-inch and 25,587 3-inch photomultiplier tubes installed. Tuning the effective parameters of these simulations to match experimental data is crucial to characterize the complex detector response and understanding detector related systematics, but traditional iterative methods are computationally prohibitive for modern, large-scale experiments like JUNO.

This work presents a novel solution using Simulation-Based Inference (SBI) to perform precise and accurate MC tuning. We achieve this by creating fast surrogate models that efficiently approximate otherwise intractable likelihoods, incorporating detector response. We developed two complementary neural likelihood estimators: (i) a Transformer Encoder Density Estimator (TEDE) for binned analysis and (ii) a Normalizing Flows Density Estimator (NFDE) suitable for both binned and unbinned analyses. Using the JUNO detector as a case study, we train these models on sets of simulated energy spectra from five distinct calibration sources, with each set generated for a specific configuration of detector response parameters. The models learn the complex, non-linear relationship between three key energy response parameters — the Birks' coefficient, the Cherenkov light yield factor, and the absolute light yield — to accurately approximate the conditional probability density function of the energy spectra for any combination of the parameters. Parameter inference is performed by integrating these learned likelihoods with a Bayesian nested sampling algorithm. Our results show that this approach successfully recovers the true parameter values with near-zero systematic bias and uncertainties limited purely by the statistics of the input data.

The proposed framework establishes a promising and generalizable template for parameter inference in modern physics experiments where a comprehensive detector response is computationally expensive to evaluate.

Speaker: Andrea Serafini

17:35 Q/A

Tuesday 28 October

09:00 → 11:35 Experiments - Cherenkov-based Neutrino Telescopes

09:00 Using End-to-End Optimization to Improve the Measurement of Neutrinos from the Galactic Plane with IceCube

Characterizing the astrophysical neutrino flux with the IceCube Neutrino Observatory traditionally relies on a binned forward-folding likelihood approach. Insufficient Monte Carlo (MC) statistics in each bin limits the granularity and dimensionality of the binning scheme. A neural network can be employed to optimize a summary statistic that serves as the input for data analysis, yielding the best possible outcomes. This end-to-end optimized summary statistic allows for the inclusion of more observables while maintaining adequate MC statistics per bin. To enable such optimization, the entire analysis pipeline must be fully differentiable. This includes differentiable binning operations and sensitivity calculation from Fisher information. Ensuring differentiability across all components allows gradients to propagate seamlessly from the final sensitivity measure back to the analysis inputs, enabling true end-to-end training. This work will detail the application of end-to-end optimized summary statistics in analyzing and characterizing the galactic neutrino flux, achieving improved resolution in the likelihood contours for selected signal parameters and models.

Speaker: Oliver Janik

09:15 Q/A

09:20 Towards Multi-task transformer based reconstruction for real-time high-energy neutrino alerts.

The IceCube Neutrino Observatory issues a variety of real-time alerts that identify high-energy neutrino events with a high probability of astrophysical origin. These alerts rely on rapid reconstruction of the incident particle's direction and energy, enabling follow-up observations by other telescopes and observatories in the context of multi-messenger astronomy. Traditionally, reconstruction methods for these alerts have been likelihood-based, requiring assumptions about event morphology. Recent advances in machine learning, both broadly and within IceCube, now enable work towards a fast and flexible event reconstruction without the need for such priors. In this talk, I will present a proposed GraphNet-based neural network approach for fast reconstruction of high-energy neutrinos.

Speaker: Aske Rosted

09:35 Q/A

09:40 Improving KM3NeT Event Reconstructions and Simulations using Generative Neural Networks

The KM3NeT collaboration is building two neutrino telescopes in the Mediterranean Sea: ORCA for low-energy oscillation studies and ARCA for the detection of high-energy astrophysical neutrinos. Both detectors are three-dimensional arrays of photomultiplier tubes that record Cherenkov light from secondary particles produced in neutrino interactions. High-level variables - such as the particle's energy, direction, and interaction point - are reconstructed using maximum-likelihood fits, which require a mapping from a set of event hypotheses, including nuisance parameters, to the expected photon-arrival-time distributions at each photomultiplier tube.

At present, this mapping uses lookup tables computed numerically from semi-analytic parameterizations. Extending the event hypothesis with additional nuisance parameters is challenging because the table’s size grows exponentially. We propose replacing the tables with a generative neural network trained on simulations that include detailed photon propagation while spanning the nuisance parameter space. The resulting model is intended to support maximum-likelihood reconstruction, fast generation of simulated events, and detector calibration. This contribution reports the project’s current status and outlines the next steps.

Speaker: Lukas Hennig

09:55 Q/A

10:00 Coffee

10:20 Machine learning at Baikal-GVD

Baikal-GVD is a large-scale underwater neutrino telescope in Lake Baikal designed to study the properties of high-energy neutrinos. In this talk, I will present the neural-network–based data processing chain currently under development for Baikal-GVD data analysis. This pipeline addresses the following goals: suppression of extensive air shower background, rejection of optical module activations caused by natural water luminescence, and reconstruction of neutrino energy and arrival direction. The developed methods improve Baikal-GVD's reconstruction accuracy and accelerate data processing. I will also discuss the challenges, including importance of domain adaptation, and outline directions for future developments.

Speaker: Ivan Kharuk

10:45 Q/A

10:55 Neural Network Approach to Neutrino and Extensive Air Shower Event Separation at Baikal-GVD experiment

This report presents a neural network–based approach for the Baikal-GVD experiment to separate neutrino-induced events from background events caused by extensive air showers (EAS). Two Transformer encoder models were developed for different stages of the data processing pipeline. The first, trained on Monte Carlo simulated raw optical module hits with spatial, temporal, and charge information, serves as a fast pre-filter. At a fixed neutrino exposure of 99%, it achieves EAS suppression factors from 2.6 (≥5 hits on two strings) up to 17 (≥16 hits on two strings). The second model, trained on noise-filtered signal hits, is designed for the final processing stages, where it preserves 64% of neutrino-induced events while achieving a background suppression factor of 10^6, making it suitable for estimating the integral neutrino flux. To improve applicability to experimental data, we consider domain-adversarial training with a gradient reversal layer, enabling the networks to learn simulation-independent features.

Speaker: Albert Matseiko

11:10 Q/A

11:15 Graph Neural Networks for Fast Simulation Reconstruction in the ESSnuSB Neutrino Detector

Graph neural networks (GNNs) have proven to excel at Cherenkov detector event reconstruction, where the signal is sparse, varying in size, and non-euclidian in structure. For the European Spallation Source Neutrino Super Beam (ESSnuSB) - which aims to measure CP-violation in the leptonic sector with an accuracy less impaired by systematic uncertainties than other proposed experiments - we have already demonstrated how GNNs can play an important role in event reconstruction in the development as well as deployment phases.
This talk recounts the main highlights of these results, and presents new studies on how GNNs can be used for fast exploration of variations in detector geometries and PMT properties in the developing stages of the ESSnuSB.

Speaker: Kaare Endrup Iversen

11:30 Q/A

13:00 → 13:55 AI/ML in HEP

13:00 AI/ML in Cosmic Frontier

Speaker: Leander Thiele

13:40 Q/A

13:55 → 16:05 Experiments - 0nuBB

13:55 Advancing Neutrinoless Double-Beta Decay in LEGEND

The observation of neutrinoless double-beta decay (0νββ) would confirm the Majorana nature of neutrinos, violate lepton number conservation, and offer deep insights into the origin of neutrino mass and the matter-antimatter asymmetry of the universe. The LEGEND (Large Enriched Germanium Experiment for Neutrinoless ββ Decay) program builds upon the success of its predecessors, the MAJORANA DEMONSTRATOR and GERDA, combining their strengths to achieve unprecedented sensitivity using enriched High-Purity Germanium (HPGe) detectors. LEGEND-200 is currently operational, while the design of the tonne-scale LEGEND-1000 is underway.
In this contribution, I will present the current status of the LEGEND experiment, with a focus on recent progress in background suppression—a critical challenge for 0νββ searches. In particular, I will highlight the role of machine learning techniques in improving background rejection, capitalizing on the detailed knowledge of signal formation in HPGe detectors. These techniques include novel, interpretable algorithms designed to meet the stringent low-background and high-transparency requirements of the experiment.

Speaker: Marta Babicz

14:20 Q/A

14:30 Machine learning for event reconstruction at SNO+

SNO+ is an operational multi-purpose neutrino detector located 2km underground at SNOLAB in Sudbury, Ontario, Canada. 780 tonnes of linear alkylbenzene-based liquid scintillator are observed by ~9300 photomultiplier tubes (PMTs) mounted outside the spherical scintillator volume. SNO+ has a broad physics program which will include a search for the neutrinoless double beta decay of 130Te.

Machine learning techniques based on transformers are being actively developed for reconstruction tasks at SNO+. We show how data from SNO+ consisting of sets of PMT hits can be effectively prepared for a transformer by tokenizing on a per-hit basis. We present transformer-based reconstruction algorithms for fitting the position and direction of events, and compare the performance of these algorithms to likelihood maximization techniques. We demonstrate position reconstruction with ML on real detector data of coincident radiogenic backgrounds, observing the same marginal improvement over likelihood maximization that is seen in MC at the same energy. We demonstrate the capability for simultaneous reconstruction of event position and direction in MC, which has proved impractical with likelihood. We show that the simultaneous fit significantly reduces bias in the position vertex for higher-energy events, while mitigating shortcomings in the direction fit associated with an imperfect position vertex.

Speaker: Cal Hewitt

14:55 Q/A

15:05 Coffee

15:25 Development of Machine Learning PID Methods in KamLAND-Zen Experiment

We developed machine learning methods for background rejection in the 0νββ decay search of the KamLAND-Zen experiment. Using CNN-based KamNet and Transformer-based ViViT for particle identification from PMT hit maps, we compared the rejection efficiency of both models. The results showed equivalent performance with high correlation (0.85-0.95) in output scores. Performance improvement through integrated models was limited.

Speaker: Jun Nakane

15:40 Q/A

15:45 Study of the performance of the NEXT-100 Tracking Plane using a Sparse Convolutional Neural Network

"The observation of neutrinoless double-beta decay (0νββ) would demonstrate that the
Neutrino is Majorana in nature (its own antiparticle),establish the violation of lepton number conservation, and provide insight into the absolute neutrino mass scale. The NEXT experiment (Neutrino Experiment with a Xenon TPC) is searching for 0νββ decay in high-pressure xenon gas at the Laboratorio Subterráneo de Canfranc (LSC) in Spain. The NEXT-100 detector can hold ∼100kg of xenon at 15 bar and consists of a high-pressure time projection chamber (HPXeTPC) using an electroluminescence (EL) region for signal amplification. The energy and the spatial pattern of ionizing particles in the detector are accurately reconstructed using two sensor planes: the Energy Plane (EP), with photomultiplier tubes (PMTs), and the Tracking Plane (TP), equipped with silicon photomultipliers (SiPMs) that collect the signal from the amplification region. The performance of the TP depends on several operational parameters such as the diffusion of ionization electrons on its way to the anode, the separation between SiPMs (SiPMs pitch) and the operational pressure of the xenon gas.

This talk will focus on the use of a Sparse Convolutional Neural Network to study the TP
performance as a function of these operational parameters with the goal of identifying the
optimal event classification capabilities of NEXT-100 under different TP configurations and
gas mixtures."

Speaker: María Cid Laso

16:00 Q/A

16:05 → 17:35 Poster

Wednesday 29 October

09:00 → 11:30 Experiments - DUNE

09:00 AI/ML at DUNE

The Deep Underground Neutrino Experiment (DUNE) is the flagship next-generation neutrino experiment in the United States, designed to decisively measure neutrino CP violation and determine the neutrino mass hierarchy. DUNE employs Liquid Argon Time Projection Chamber (LArTPC) technology, which provides exceptional spatial resolution and enables detailed reconstruction of final-state particles and neutrino interactions. However, this high granularity demands advanced and cutting edge reconstruction techniques to unlock its full potential.

Artificial intelligence and machine learning (AI/ML) techniques—such as convolutional neural networks, graph neural networks, and transformers—are being actively developed within DUNE and have already demonstrated strong performance in signal processing, kinematic reconstruction, clustering, and interaction/particle identification. Beyond reconstruction, AI/ML methods are playing an increasingly important role in simulation, trigger and DAQ data processing, beam design and monitoring, documentation search, and quality assurance/quality control (QA/QC). In parallel, DUNE has been developing dedicated AI/ML infrastructures to support these efforts.

In this talk, I will review the progress of AI/ML approaches and the supporting facilities within DUNE.

Speaker: Jianming Bian

09:25 Q/A

09:35 Capturing Neutrinos and New Physics at ProtoDUNE with a Machine Learning-Based Trigger Algorithm

The Deep Underground Neutrino Experiment (DUNE) will deploy massive liquid argon time projection chamber (LArTPC) detectors to measure the properties of neutrinos with unprecedented precision. Currently, two full-scale prototypes of the DUNE far detectors (FDs) are at CERN, which are called the ProtoDUNE detectors. Recent studies indicate that these prototypes could have the potential to detect neutrinos and long-lived beyond-standard model (BSM) particles from one of the targets in CERN’s north area that is exposed to the 400 GeV Super Proton Synchrotron (SPS) beam. A limiting factor in any neutrino and BSM physics program at ProtoDUNE is the ability to trigger for the interesting signal whilst rejecting the overwhelming cosmic-ray background present at these surface detectors. A new exclusive self-triggering algorithm has been developed for the DUNE data acquisition (DAQ) system that uses an XGBoost gradient boosted decision tree (GBDT) algorithm to select data with neutrino-like properties and reject cosmic-ray background. Simulations demonstrate that this algorithm improves the efficiency of selecting neutrino events at ProtoDUNE compared to existing algorithms. This talk will explain how such an ML-based triggering algorithm can enhance a future neutrino and BSM physics program at ProtoDUNE.

Speaker: Ciaran Hasnip

09:50 Q/A

09:55 Deep Neural Network Cascade for the Deep Underground Neutrino Experiment

"Particle imaging detectors have been central to particle physics for more than a century, providing an unrivaled level of detail that has enabled numerous discoveries. Liquid argon time projection chambers (LArTPCs), a dense and scalable realization of this detection paradigm, constitute the core technology of the Deep Underground Neutrino Experiment (DUNE). Automating the reconstruction of particle interactions in LArTPCs has remained a major challenge; without reliable solutions, the physics program of DUNE could be significantly compromised.

Recent advances in machine learning (ML), particularly in computer vision, offer a path forward. We present the Scalable Particle Imaging with Neural Embeddings (SPINE) framework: a machine-learning–based reconstruction chain for particle imaging detectors. SPINE employs a multi-task neural network cascade that integrates voxel-level feature extraction via sparse convolutional neural networks with particle superstructure construction via graph neural networks. This approach enables detailed characterization of neutrino interactions and is currently deployed in three experiments for state-of-the-art physics inference.

The DUNE near detector will operate under unprecedented conditions, including pile-up of O(100) neutrino interactions in a O(100) ton detector and the highest neutrino energies yet encountered in a long-baseline experiment. We report on the first application of SPINE to this environment and provide a detailed evaluation of its performance. Prospects for further development, including extensions targeting hadronic shower energy reconstruction and particle-flow reconstruction, are also discussed."

Speaker: Francois Drielsma

10:20 Q/A

10:30 Coffee

10:50 Track Matching Across Detectors: Using GNNs to Match Particles across DUNE’s Near Detector Prototypes

In the global scientific effort to better understand how neutrinos fit (or don’t) within the bounds of the Standard Model, the Deep Underground Neutrino Experiment (DUNE) aims to make precise neutrino oscillation measurements to determine the neutrino mass ordering and establish the value of neutrino Charge-Parity (CP) violation. To accomplish this, DUNE has a host of near detectors that will be placed next to the source of the world’s most intense accelerator neutrino beam in order to characterize the neutrino interactions and to constrain measurements performed 1300 km away at the far detectors’ site. To capture particles leaving the Liquid Argon (LAr) near detector volume, with a particular focus on muons, a muon tagger is placed downstream. Precisely matching the particles across the detectors during the reconstruction phase can help improve the final particle ID determination and help us cope with the very large pile-up expected in the intense neutrino beam. This work shows the potential of using Graph Neural Networks to connect track segments between the solid scintillator detector planes to the central LAr detector region. This is being developed using the current setup of DUNE’s prototype LAr near detector, the “2x2,” and the solid scintillator muon tagger provided by repurposed MINERvA planes.

Speaker: Jessie Micallef

11:05 Q/A

11:10 DNN + MLEM synergy for imaging of neutrino interactions in LAr

DUNE is a next-generation long-baseline neutrino experiment aiming to determine
the neutrino mass ordering, study CP violation in the leptonic sector, observe
supernova neutrinos, and search for physics beyond the Standard Model. It will
feature a Near Detector 547 m from the source and a Far Detector ~1300 km away.
Within the Near Detector, the System for on Axis Neutrino Detection includes GRAIN
(GRanular Argon for Interactions of Neutrinos), a novel liquid argon detector
designed to image neutrino interactions via scintillation light, providing vertexing and
tracking.
GRAIN features an innovative cryogenic light readout system consisting of a matrix
of SiPMs with optics based on coded aperture masks (grids of alternating opaque
material and holes). The reconstruction algorithm, based on Maximum Likelihood
Expectation-Maximization (MLEM), combines the views of ~60 cameras providing a
three-dimensional map of the energy deposited by charged particles. This iterative
approach presents a significant computational challenge, requiring optimized use of
multiple GPUs.
The aim of this work is to provide a prior of the expected three-dimensional energy
deposition to serve as a seed for the MLEM algorithm, rather than a uniform
distribution. This improves convergence and reduces GPU load. A deep neural
network (DNN) was trained on simulated charged-current muon neutrino
interactions, using timing information from the cameras. Efficient hyperparameter
optimization was carried out using the OPTUNA framework. The resulting model
produces a better seed for the MLEM algorithm.
The synergy between machine learning and classical algorithms leverages the
speed of DNN predictions together with the precision of MLEM.

Speaker: Filippo Mei

11:25 Q/A

11:30 → 11:50 Neural Inference Techniques

11:30 Bayesian Neural Networks for classification problems with LAr TPCs

Liquid Argon Time Projection Chambers (LAr TPCs) provide detailed imaging of neutrino interactions, making them an ideal environment for machine learning based classification tasks. While conventional neural networks have achieved strong performance in this domain, they do not naturally account for uncertainties, an essential requirement for robust physics analyses. Bayesian Neural Networks (BNNs) address this limitation by treating network parameters probabilistically, enabling classification outputs with calibrated uncertainty estimates. In this talk, we present applications of BNNs to classification problems in neutrino experiments with LAr TPCs.

Speaker: Aaron Higuera

11:45 Q/A

13:35 → 14:10 Experiments - NOvA

13:35 Heterogeneous Point Set Transformers for Segmentation of Multiple View Detectors

NOvA is a long-baseline neutrino experiment studying neutrino oscillations with Fermilab's NuMI beam. Prong reconstruction, the task of matching detector hits to their source particles and identifying the type of each particle, is a crucial and resource-intensive step in the event reconstruction process. This task has commonly been done using traditional clustering approaches or convolutional neural networks (CNNs). Due to the construction of the detector, the data is presented as two sparse 2D images: an XZ and a YZ view of the detector, rather than a 3D representation. We propose a point set neural network that operates on the sparse matrices with an operation that mixes information from both views. Our model uses less than 10% of the memory required using previous methods while achieving higher segmentation accuracy compared to when both views are processed independently.

Speaker: Alejandro Yankelevich

14:00 Q/A

14:10 → 16:25 Experiments - SBN

14:10 End-to-End, Machine-Learning-Based Event Reconstruction and Detector Calibration in ICARUS

The ICARUS detector, a LArTPC (Liquid Argon Time Projection Chamber) of 476 tons fiducial volume, serves as the Far Detector of the SBN (Short Baseline Neutrino) program. ICARUS is situated on-axis with respect to the BNB and off-axis to the NuMI neutrino beams at Fermilab. LArTPC is a powerful detector technology for achieving precise neutrino interaction imaging and reconstruction in 3D, thanks to its mm-scale spatial resolution. An end-to-end, scalable reconstruction chain referred to as "SPINE" (Scalable Particle Imaging with Neutral Embeddings) makes the use of Sparse Convolutional Neural Network on the voxel-level information extraction and Graph Neural Network on the particle-level structure clustering in a hierarchical way to analyze data. This talk presents the reconstruction performance of SPINE on two well-known processes in ICARUS: Michel electron and neutral pions decaying to two photons. Michel electrons are the daughter particles of muon decay-at-rest that have well-defined energy spectra, making them suitable targets of detector energy scale calibration below 100 MeV; meanwhile, the neutral pion decaying into a pair of gamma photons is one of the dominant background in the ICARUS electron neutrino appearance measurement due to the presence of electromagnetic showers. Understanding both processes in ICARUS is key to the successful application of machine learning techniques toward SBN neutrino oscillation physics.

Speaker: Junjie Xia

14:35 Q/A

14:45 Reconstruction of Electron Neutrino Interactions at ICARUS with SPINE Machine Learning

The Short-Baseline Neutrino (SBN) Program at Fermilab consists of two liquid argon time projection chamber (LArTPC) detectors: the Short-Baseline Near Detector (SBND) located 110 meters downstream from the Booster Neutrino Beam (BNB) target, and the ICARUS detector positioned 600 meters downstream from the BNB target. The program is designed to probe short-baseline neutrino anomalies, including the LSND electron neutrino excess and the MiniBooNE low-energy excess. Additionally, the ICARUS detector, which also lies off-axis to the NuMI beamline, provides unique sensitivity to physics beyond the Standard Model (BSM) and novel cross section measurements not accessible through the BNB alone. To analyze the data from these detectors, we have begun employing a machine-learning-based reconstruction algorithm referred to as “Scalable Particle Imaging with Neural Embeddings” (SPINE). SPINE has shown improvement in neutrino identification and particle species discrimination for both track-like and shower-like topologies in ICARUS, with the potential to enhance the quality of measurements achievable within the SBN Program. In this talk, I will present an overview of electron neutrino analysis efforts at ICARUS that utilize SPINE, highlighting its impact on particle selection and reconstruction performance.

Speaker: Daniel Carber

15:00 Q/A

15:05 Single Photon Searches at ICARUS with SPINE

The MiniBooNE Low-Energy Excess (LEE) of electron-like events from the Booster Neutrino Beam (BNB) has puzzled neutrino physicists for decades. One possible explanation has been an unpredicted excess of neutral current (NC) ∆ resonance interactions with a subsequent radiative decay. An increase in the rate of NC ∆→Nγ events by a factor of 3.18 could explain the LEE seen by MiniBooNE. ICARUS also sees neutrinos from the BNB and can check the rate of these single photon events. The SPINE particle physics reconstruction suite leverages deep neural networks (DNNs) to optimize particle reconstruction and identification in Liquid Argon Time Projection Chambers (LArTPCs). I present preliminary findings on the effectiveness of these machine learnign (ML) techniques for single photon event reconstruction at ICARUS.

Speaker: Harry Hausner

15:20 Q/A

15:25 Coffee

15:45 Performance of Muon Neutrino Reconstruction at ICARUS using SPINE Machine Learning

ICARUS is a liquid argon time projection chamber (LArTPC) neutrino experiment at Fermilab. Located ~600 m from the Booster Neutrino Beam (BNB) target, it serves as the far detector in the Short-Baseline Neutrino (SBN) Program. The primary objective of the SBN Program is to probe neutrino oscillation physics at short baselines using multiple detectors, including measurements of both muon neutrino disappearance and electron neutrino appearance. A key requirement for the neutrino oscillation measurements is robust event reconstruction, including accurate and precise particle identification (PID) and kinematic reconstruction across a wide range of topologies. To achieve the required performance, we employ the SPINE ("Scalable Particle Imaging with Neural Embeddings") package, a state-of-the-art machine learning reconstruction framework. This talk presents the status of muon neutrino reconstruction using SPINE at ICARUS, including the evaluation of reconstruction performance that will enable the full physics program at ICARUS and at the SBN Program.

Speaker: Dante Totani

16:00 Q/A

16:05 End-to-End Machine Learning Reconstruction for the Short Baseline Near Detector

The Short-Baseline Near Detector (SBND) is a 100-ton scale Liquid Argon Time Projection Chamber (LArTPC) neutrino detector positioned in the Booster Neutrino Beam (BNB) at Fermilab, as part of the Short-Baseline Neutrino (SBN) program. Recent inroads in Computer Vision (CV) and Machine Learning (ML) have motivated a new approach to the analysis of particle imaging detector data. SBND data can therefore be reconstructed using an end-to-end, ML-based data reconstruction chain for LArTPCs. The scalable particle imaging with neural embeddings (SPINE) reconstruction chain is a multi-task network cascade which combines point-level feature extraction using Sparse Convolutional Neural Networks (CNN) and particle superstructure formation using Graph Neural Networks (GNN). SPINE has been trained using a particle bomb simulation that is propagated through SBND's detector simulation suite. This talk will demonstrate neutrino selections that utilize SPINE at SBND.

Speaker: Bear Carlson

16:20 Q/A

16:25 → 17:50 Poster

Thursday 30 October

09:00 → 10:50 Foundation Models

09:00 PID Without Labels: Toward a Foundation Model for All Neutrino Experiment Reconstruction Tasks

We present what to our knowledge is the first sensor-level foundation model (FM) for neutrino detectors, trained directly on simulated 3D LArTPC charge data without manual labels. The goal is simple: build a scalable model that learns the underlying physics automatically from data, then adapts to event reconstruction, PID, calibration, and other tasks using only a small labelled sample. I will discuss two complementary self-supervised approaches. The first is masked autoencoding, which forces the model to internalize detector response and interaction structure by reconstructing randomly masked portions of particle trajectories. The second is self-distillation, a teacher-student training setup that builds robust representational invariances with data-corrupting augmentations. We share what makes these approaches work well with LArTPC data, and show that on downstream tasks, our model delivers large sample-efficiency gains; for example, fine-tuning on just 100 example images results in >98% F1 score in segmenting tracks and showers. The net effect is an extensible physics-aware backbone for LArTPC data.

Speaker: Samuel Young

09:25 Q/A

09:35 Conditional Generation of LArTPC Images Using Latent Diffusion

Given the challenges in LArTPC event reconstruction, we have taken the first steps towards creating an end-to-end inference pipeline to go from 2D images to event properties. Inspired by the success of denoising diffusion probabilistic models (DDPMs), we have developed a method of conditionally generating 2D LArTPC images. By utilizing a modified latent diffusion model, we have demonstrated the ability to generate single-particle events of a specified momentum with quality comparable to traditional Geant4 simulations.

Speaker: Zev Imani

10:00 Q/A

10:10 Coffee

10:30 Advancing Event Reconstruction Techniques for High-Energy Neutrino Physics Using Conventional and Deep Learning Approaches for the FASER Experiment at the CERN LHC

The FASERCal detector, proposed as an off-axis neutrino detector for the FASER experiment to operate during LHC Run 4, will produce sparse, 3D voxelized data, demanding advanced deep learning for neutrino event reconstruction. We present a hybrid architecture that uses a Sparse Submanifold Convolutional Network (SSCN) to efficiently tokenize voxel hits into patch embeddings. These are processed by a hierarchical Transformer encoder with attention mechanisms designed to capture features at multiple spatial scales, learning both fine-grained local shower structures and the global event topology. The model is trained using a two-stage transfer learning approach. A self-supervised pre-training phase employs a dual-objective Masked Autoencoder (MAE) to learn a robust physical representation via both generative reconstruction and contrastive learning. The resulting encoder is then fine-tuned for multi-task classification and kinematic regression using task-specific cross-attention heads. This framework achieves high-fidelity reconstruction on simulated data, showing stable performance and first signs of sensitivity to $\nu_{\tau}$ CC interactions and charm production in neutrino events.

Speaker: Fabio Cufino

10:45 Q/A

10:50 → 11:30 Innovative PID Approaches

10:50 Proton/Deuteron Separation at the LArIAT Experiment

We present a study of proton–deuteron separation in the Liquid Argon In A Testbeam (LArIAT) experiment at Fermilab. LArIAT, which operated from 2015-2017, was designed to study and characterise charged particles in liquid argon that commonly occur in neutrino-Ar interaction final-states. Deuterons are rare and subject to significant background contamination in the LArIAT dataset, and their short, highly ionising tracks can closely resemble those of stopping protons, making them difficult to distinguish. Yet, this separation is crucial in reconstructing neutrino events accurately. Here, we present techniques, machine learning and otherwise, to separate proton and deuteron tracks and discuss their performance and limitations.

Speaker: Mohammed Sultan

11:05 Q/A

11:10 Optimal Transport for $e/\pi^0$ Particle Classification in LArTPC Neutrino Experiments

The efficient classification of electromagnetic activity from $\pi^0$ and electrons is a notoriously challenging problem in the reconstruction of neutrino interactions in Liquid Argon Time Projection Chamber (LArTPC) detectors. We address this problem using the mathematical framework of Optimal Transport (OT), which has been successfully employed for event classification in other HEP contexts and is ideally suited to the high-resolution calorimetry of LArTPCs. Using a publicly available simulated dataset from the MicroBooNE collaboration, we show that OT methods achieve state-of-the-art reconstruction performance in $e/\pi^0$ classification. This performance is further improved when we couple the OT outputs to interpretable machine learning methods, including k-Nearest Neighbors (kNN) and Support Vector Machine (SVM). The success of this first application indicates the broader promise of OT methods for LArTPC-based neutrino experiments such as SBN and DUNE. Since $\pi^0$s are a significant background for both oscillation experiments and BSM searches, integrating OT can lead to sizeable improvements in the selection efficiency for these analyses by introducing a novel method with which to achieve $\pi^0$ rejection.

Speaker: Chuyue “Michaelia” Fang

11:25 Q/A

13:00 → 15:45 AI//ML for Detector Physics Modeling

13:00 Determining a TPC Electric Field using Physics-Informed Continuous Normalizing Flows

Noble element time projection chambers (TPCs) have enabled searches for neutrino physics both within and beyond the Standard Model, but these rare event searches are limited by electric field distortions on event position reconstruction. Traditional data-driven field distortion corrections lack smoothness and differentiability, require extensive calibration data, and often do not follow Maxwell's equations. We present a physics-informed continuous normalizing flow approach that models the electric field lines by learning an invertible, differentiable, curl-free transformation from reconstructed positions to true interaction vertices. We evaluate this machine learning method's performance against traditional field distortion correction methods. By improving event position reconstruction, future rare event searches can increase the number of correctly reconstructed interaction events and achieve greater wall background rejection.

Speaker: Ivy Li

13:25 Q/A

13:35 PMT Timing Calibration in Large Liquid Scintillator Detectors with In Situ Radioactive Backgrounds

Event reconstruction in large liquid scintillator neutrino detectors, such as SNO+, rely on hit times from large numbers of photomultiplier tubes (PMTs). Standard methods of PMT timing calibrations involve dedicated hardware with deployed or in situ light sources. A calibration using in situ radioactive backgrounds present in regular physics data would allow vastly more frequent calibrations without the use of dedicated calibration hardware or the risk of radioactive contamination from deployed sources. We present a novel method that uses a basic scintillator emission timing distribution to simultaneously train a position reconstruction neural network and a simple PMT timing calibration model on radioactive backgrounds. We show that this calibration method applied to SNO+ data is comparable to a standard calibration.

Speaker: Scott DeGraw

13:50 Q/A

13:55 LUCiD: a Light-based Unified Calibration and trackIng Differentiable simulation

Next-generation monolithic Water Cherenkov detectors aim to probe fundamental questions in neutrino physics. These measurements demand unprecedented precision in detector calibration and event reconstruction, pushing beyond the capabilities of traditional techniques. We present a novel framework for differentiable simulation of Water Cherenkov detectors that enables end-to-end optimization through gradient-based methods. By leveraging JAX's automatic differentiation and implementing a grid-based acceleration system, our framework achieves millisecond-scale simulation times - four orders of magnitude faster than traditional approaches. The framework can incorporate neural network surrogates for unknown physical phenomena while maintaining interpretability throughout the simulation chain. As a demonstration, we employ a neural network to model differentiable photon generation probability distributions. Our modular architecture extends to various Water Cherenkov detectors, representing a significant step toward addressing systematic limitations in future neutrino experiments through differentiable programming techniques.

Speaker: Omar Alterkait

14:20 Q/A

14:30 Uncertainty propagation with a LArTPC differentiable simulator

Liquid argon time projection chambers (LArTPCs) are highly attractive for particle detection because of their tracking resolution and calorimetric reconstruction capabilities. Developing high-quality simulators for such detectors is very challenging because conventional approaches to describe different detector parameters or processes ignore their entanglement (ie, calibrations are done one at a time), which translates into a poor description of the underlying physics by the simulator. To address this, we created a differentiable simulator that enables gradient-based optimization, allowing an in-situ simultaneous calibration of all detector parameters for the first time. The simulator has been demonstrated to robustly fit targets across a wide range of parameter space using multiple physics samples, and therefore provides a strong proof-of-concept demonstration of the utility of differentiable detector simulation for the calibration task. In this talk, I will focus on the possibilities of using such a simulator to perform uncertainty propagation from the physical parameters down to the final reconstructed quantities.

Speaker: Pierre Granger

14:55 Q/A

15:05 Towards data application of simultaneously calibrating multiple detector effects using differentiable simulation

Missing correlations and potential biases in detector calibrations are one major challenge mitigating data—simulation differences. To address this, we propose to use gradients, enabled by differentiable simulation, for efficient and effective optimization of the detector physics parameters. We developed an auto-differentiation enabled simulation of a liquid argon time projection chamber using JAX, and applied it on calibrating multiple detector effects simultaneously. This approach allows us to account for the correlations of the detector modeling parameters comprehensively and avoid biases introduced by segmented measurements. In this talk, I will present detector calibration using data-like samples and discuss practical considerations for deploying this method in experimental settings.

Speaker: Yifan Chen

15:20 Q/A

15:25 Coffee

15:45 → 17:50 New Experiments and Datasets

15:45 Nuclear recoil detection with color centers in bulk lithium fluoride

We present initial results on nuclear recoil detection based on the fluorescence of color centers created by nuclear recoils in lithium fluoride. We use gamma rays, fast and thermal neutrons, and study the difference in responses they induce, showing that this type of detector is rather insensitive to gamma rays. We use light-sheet fluorescence microscopy to image nuclear recoil tracks from fast and thermal neutron interactions deep inside a cubic-centimeter sized crystal and demonstrate automated feature extraction in three dimensions using machine learning tools. The number, size, and topology of the events agree with expectations based on simulations with TRIM. These results constitute the first step towards 10-1000g scale detectors with single-event sensitivity for applications such as the detection of dark matter particles, reactor neutrinos, and neutrons.

Speaker: Patrick Stengel

16:10 Q/A

16:20 3D particle tracking in novel unsegmented scintillators

The full abstract is not included at this stage as the work is currently under patent submission. The complete abstract will be made available once the patent has been published, which should be before the conference.

Speaker: Saul Alonso-Monsalve

16:45 Q/A

16:55 Event Classification and Annotated Dataset in CONNIE

The COherent Neutrino–Nucleus Interaction Experiment (CONNIE), located 30 m from the core of the Angra 2 reactor at the Almirante Álvaro Alberto Nuclear Power Plant in Brazil, is the first reactor neutrino experiment to employ silicon Skipper-CCDs with the aim of detecting coherent elastic neutrino-nucleus scattering (CEvNS), searching for physics beyond the Standard Model, and monitoring reactor activity. A key challenge is distinguishing between various event types within large volumes of unlabeled data, and to address this, machine learning methods are applied to classify CONNIE events, providing a benchmarking analysis. For this purpose, a labeled dataset of real experimental events is developed using the Annotation Redundancy with Targeted Quality Assurance method, in which multiple annotators independently label the same subset of data to identify and resolve discrepancies, building a reliable and robust dataset. We will present preliminary results demonstrating the performance of the classification models and the progress in constructing the annotated dataset.

Speaker: Sara Mirthis Dantas dos Santos

17:10 Q/A

17:15 Data Challenges for Neutrino Physics AI research

TBD

Speaker: Kazuhiro Terao

17:40 Q/A

Friday 31 October

09:00 → 10:20 AI/ML for New Physics Search

09:00 Search for keV-Sterile Neutrinos with TRISTAN at KATRIN Using a Neural Network-Based Approach

Following the completion of its neutrino mass measurement program at the end of 2025, the KATRIN experiment aims to probe keV-scale sterile neutrinos by analyzing the full tritium beta decay spectrum with a novel detector system, TRISTAN. Leveraging KATRIN’s high source activity, this search is sensitive to mixing amplitudes at the parts-per-million level. However, extracting a potential sterile neutrino signature is challenging, as it relies on detailed modeling of the the observed tritium spectrum and requires computationally intensive Monte Carlo simulations. To address this challenge, we explore a neural network-based approach to identify sterile neutrino signatures directly from the spectral data. In this talk, we will present the expected sensitivity of this method and evaluate its robustness against key modeling uncertainties. In addition, we demonstrate how flow-matching techniques can be leveraged to overcome limitations of traditional Monte Carlo simulations, enabling more efficient modeling of the experimental spectra.

Speaker: Luca Fallböhmer

09:15 Q/A

09:20 Real-time Anomaly Detection in Liquid Argon Time Projection Chamber

Real-time anomaly detection provides a model-independent way to search for unexpected phenomena, complementing traditional model-driven triggers. At the LHC, autoencoder-based anomaly triggers in CMS such as CICADA (on raw calorimeter data) and AXOL1TL (on reconstructed objects) have already demonstrated the power of these methods, with similar efforts underway in ATLAS through GELATO. Motivated by these successes, we investigate the application of such techniques to Liquid Argon Time Projection Chambers (LArTPCs), where raw wire data directly capture ionization energy depositions. I will present results on the physics performance of autoencoder-based anomaly detection networks for LArTPCs, along with initial benchmarking studies toward hardware acceleration on FPGAs and CPUs.

Speaker: Seokju Chung

09:35 Q/A

09:40 Unsupervised clustering for neutrino event exploration

Neutrino interactions are poorly understood and modeled, with some processes entirely missing from simulation. This sometimes leads to to surprising observations from rudimentary data analysis techniques such as hand-scanning. Here, the use of contrastive learning and unsupervised clustering is investigated as a more systematic way to explore event types in data. Although the aim is to investigate unmodeled neutrino interaction mechanisms, this describes a pilot study with cosmic muons as a proof of principle.

Speaker: Callum Wilkinson

09:55 Q/A

10:00 Coffee

10:20 → 11:20 Neutrino-Nucleus Interaction Modeling

10:20 Neutrino-Nucleus Interactions from Deep Learning

In this talk, I will review recent work from our group on developing deep learning models for lepton–nucleus interactions. I will begin by applying generative adversarial networks (GANs) to simulate neutrino and antineutrino collisions with nucleons. Our models encompass both charged-current quasielastic and inclusive interactions of muon neutrinos with a carbon target, offering detailed predictions for the final charged lepton. I will then discuss how transfer learning can extend knowledge from one scattering process to another. For example, I will show that a deep neural network for cross-sections, pre-trained on electron–carbon scattering data, can accurately reproduce electron scattering in helium-3, lithium, oxygen, calcium, aluminum, and iron after only minimal fine-tuning with limited new measurements. Finally, I will present results on applying domain adaptation to GAN models that generate neutrino and antineutrino interactions with nuclei. This talk is based on the following references: Phys. Rev. Lett. 135, 052502; Phys. Rev. D 112 (2025) 1, 013007; and arXiv:2508.12987.

Speaker: Krzysztof Graczyk

10:45 Q/A

10:55 Surrogate model for final-state interactions of neutrino generator

Final-state interactions (FSI) are a significant source of systematic uncertainty in neutrino event generators, particularly for LArTPC-based experiments. These interactions occur when particles produced in the initial neutrino-nucleus collision scatter or are absorbed while exiting the nucleus, complicating event reconstruction and interpretation. While advanced models such as GiBUU offer accurate simulations of FSI using many-body quantum transport theory, their high computational cost limits their applicability in large-scale analyses. In this talk, I will present ongoing work on developing a machine learning-based surrogate model designed to emulate the behavior of more detailed FSI simulations. I will discuss the current status, effectiveness, and the challenges and limitations associated with this approach.

Speaker: Patrick Tsang

11:10 Q/A

13:00 → 13:55 Neural Inference Techniques

13:00 Machine Learning Assisted Reweighting and Unfolding for Neutrino Analyses

It is well-known that classifiers can be trained to approximate the likelihood ratio between two distributions, and that machine learning based classifiers in particular can learn this likelihood ratio in high-dimensional spaces. This provides a method to reweight events from different distributions as functions of many features. OmniFold is an unfolding technique that uses this concept to perform unbinned, high-dimensional unfolding. This talk will present the application of OmniFold to a neutrino cross-section study using T2K public data and discuss potential future applications to other neutrino experiments.

Speaker: Roger Huang

13:25 Q/A

13:35 Treating Detector Systematics via a Likelihood Free Inference Method

Quantifying systematic uncertainties in particle physics analyses is complicated by the need to estimate detector responses through Monte Carlo (MC) simulations. Conventional approaches compute variations in reconstructed variables under fixed physics assumptions, which can bias the parameters being measured. We present a likelihood-free inference method that uses neural networks to derive event-wise reweighting factors from MC simulations of varied detector realizations. These weights describe how detector response changes with detector properties, independent of any physics model. Applied to a simplified neutrino oscillation experiment, the method cleanly separates detector modeling from physics parameter estimation in MC forward-folding analyses.

Speaker: Alexandra Trettin

13:50 Q/A

13:55 → 14:50 Conference Organization

13:55 NPML 2026 and forward

Speaker: Kazuhiro Terao

14:20 Q/A

14:30 Closing

Speakers: Kazuhiro Terao, Patrick de Perio (Kavli IPMU)