Self-Supervised Learning Sharpens Resonance Mass

A new machine learning technique promises to sharpen the search for new physics at particle colliders. Researchers have developed a self-supervised learning model that more accurately reconstructs the mass of heavy, hypothetical particles from their complex decay signatures.

Reconstructing the mass of a heavy resonance from its decay products is a central task in new physics searches. The process is complicated by missing energy and various systematic uncertainties that can distort results. Supervised learning models, which are trained on labeled data, often struggle to generalize when faced with these distortions. Exhausting all possible variations in training data is computationally expensive, and a model's failure can corrupt the reconstructed resonance widths critical for identifying peaks in data.

The Self-Supervised Approach

To address this, the team adopted a foundation model paradigm. They first pre-trained a transformer encoder using a self-supervised objective called VICReg. This pre-training phase aimed to learn an embedding, or internal representation, that is invariant to various data corruptions. The model was then fine-tuned specifically for the task of mass regression. The study focused on a heavy resonance with masses ranging from 2.5 to 6.5 teraelectronvolts (TeV) decaying via a SUSY-like cascade into an eleven-body final state.

Performance Against Corruptions

The results showed a clear advantage for the self-supervised method. The pre-trained model reconstructed sharper resonance peaks in the mass spectrum. More importantly, it demonstrated more stable performance under various realistic corruptions when compared to a supervised model of identical architecture. The supervised model was trained from scratch on the same augmented data but did not match the robustness of the pre-trained version.

This work highlights a shift towards more strong machine learning in high-energy physics. The systematic procedure of an experiment involves controlling conditions and analyzing results to test a hypothesis. A model that fails under shifted conditions can jeopardize that empirical evidence. The self-supervised approach, by learning invariant features beforehand, acts as a more reliable component in the experimental chain.

The research was detailed in a paper submitted to the arXiv preprint server on September 15, 2026. The authors conclude that their method offers a more compute-efficient path to building models resilient to the systematic uncertainties endemic to collider data analysis.