RODEM

Robust Deep Density Models for
High-Energy Physics and Solar Physics

High Energy Physics and Solar Physics face similar challenges.

Finding relevant cases in this data deluge

can potentially revolutionize physics!

But we need to find the needle in the haystack.

RODEM was a SNSF/SINERGIA project fostering cooperation between high energy physicists, solar physicists and experts in machine learning in Switzerland to advance research methodologies in both fields. Research activities took place between Dec 2020 and April 2025.

The starting point

During the last decades, the amount of data available to scientists has increased enormously. New infrastructures, such as the Large Hadron Collider (LHC), and a new generation of solar observatories, such as the Solar Dynamics Observatory (SDO), produce data on a scale that could not be exploited to their full extent using the methods available at the beginning of the project.

Simultaneously, data science has experienced real game changing breakthroughs in the past years. In particular, deep learning methods have shown the potential of data driven approaches compared to traditional algorithmic approaches.

The goals

RODEM aimed to develop large-scale machine learning methods with provable guarantees for rare event detection and prediction in High Energy Physics (HEP) and solar astronomy. The core objectives were to create:

1. improved forecasting tools trainable from large, high-dimensional datasets with limited supervision
2. computationally efficient generative models as surrogates for expensive simulators
3. anomaly detectors with formal guarantees for super-rare event regimes

Areas in which we achieved results include

Anomaly Detection: Here, research produced a complete methodological transformation from ‘mass sculpting’ to production deployment, culminating in the New Physics Learning Machine. NPLM is now used within the ATLAS collaboration for active new physics searches.

Generative Modeling: Methodological advances lead to a complete pipeline from individual particle clouds to full event generation, representing the first transformer-coupled diffusion model for variable-multiplicity events—crucial for High-Luminosity LHC simulation.

Theoretical Foundations and Computational Innovations: The theoretical foundations developed through this project contribute to the broader machine learning literature while solving specific problems in physics simulations. Computational efficiency innovations speed up ML processes by a factor of 100, enabling training of generative models and addressing fundamental scalability bottlenecks in large scale data processing. These techniques have broad applicability in the current landscape of large language models and foundation model research.

We successfully adapted our methods for cross-domain extensions to Gaia astrometry, validating method robustness and generalisability. Our findings have the potential to improve automated tools for monitoring space weather and contribute to a deeper understanding of the physical processes driving space weather phenomena

The project has resulted in over 25 peer-reviewed publications in top-tier venues including multiple NeurIPS conferences, with strong impact across both machine learning and physics communities. All major methodological contributions have been released as open-source software with comprehensive documentation, enabling community adoption and further development.

For a full list and detailed descriptions of work performed and results achieved please consult the scientific report.