RINO: Renormalization Group Invariance with No Labels

Zichun Hao, Raghav Kansal, Abhijith Gandrakota, Chang Sun, Ngadiuba Jennifer, Javier Duarte, Maria Spiropulu

TL;DR

RINO (Renormalization Group Invariance with No Labels) is a new self-supervised learning framework designed for high-energy physics that pretrains models directly on unlabeled particle collision data. By using physics-informed jet clustering as a data augmentation strategy, RINO learns representations invariant to energy scales, mitigating the classic simulation-to-data domain shift and boosting cross-domain jet classification accuracy by up to 26%.

What Problem Does This Paper Address?

In high-energy physics (HEP), training machine learning models typically requires massive, perfectly labeled datasets. Because we cannot look at a real particle collision at the Large Hadron Collider (LHC) and instantly know its exact quantum history, physicists rely heavily on Monte Carlo (MC) simulations to generate labeled data.

However, simulations are imperfect. They often mismodel complex physical processes (like parton showering and hadronization) or subtle detector responses. This creates a domain shift: a model trained to perfection on simulation often performs poorly when deployed on real experimental data. Meanwhile, petabytes of real collision data go completely unused during training simply because they lack labels. RINO addresses this gap by introducing a way to leverage unlabeled data for pretraining.

Methods

RINO adapts the popular DINO self-supervised framework—which relies on teacher-student self-distillation—to the unique constraints of particle physics.

Standard computer vision augmentations like cropping, resizing, or color jittering fail for particle physics data because a jet's constituents possess highly global correlations. Instead, the authors introduce a physics-informed augmentation based on the $k_T$ clustering algorithm. This algorithm iteratively groups particles together, effectively mimicking the reverse process of a parton shower.

Figure 1: RINO training strategy (left): pretrain on JETCLASS QCD jets as a proxy for real data, then finetune on JETNET top tagging. Supervised baseline strategy (right): train from scratch on JETNET. Both strategies are then evaluated on top tagging on JETCLASS.

By stopping the clustering at different depths, RINO generates multiple "views" of the same jet at different energy scales (corresponding to the quantum chromodynamic, or QCD, renormalization group flow):

  • Global Views: Coarse-grained descriptions containing fewer subjets (1 to 4), given to both the teacher and student networks.
  • Local Views: Fine-grained descriptions containing more subjets (8 to 64) plus the original particle-level data, given only to the student network.

This setup forces a Transformer encoder backbone to learn representations that are invariant to the energy scale. For downstream tasks, the authors utilize a two-stage LP-FT (Linear Probing then Fine-Tuning) strategy, which initially freezes the backbone to train the classification head before unfreezing the entire network with a reduced learning rate to prevent overfitting.

Datasets

The authors emulate the simulation-to-data domain shift using two distinct jet simulation frameworks:

  • JETCLASS: Used as a proxy for real, unlabeled experimental data. The model is pretrained exclusively on the Quantum Chromodynamic (QCD) background jets from this dataset, mimicking the QCD-dominated nature of real collider data.
  • JETNET: Used as the smaller, high-quality labeled simulation dataset for the downstream task of top tagging (discriminating between background QCD jets and jets originating from top quark decays).

Key Insights and Ideas

  • Physics as an Inductive Bias: Instead of forcing a model to be invariant to arbitrary geometric distortions (like image rotations), RINO forces invariance to a real physical concept: the energy scale of the physical process.
  • Learning from the Background: Remarkably, even though RINO is pretrained only on background QCD jets, it learns rich, underlying physical features that make it highly effective at identifying top quark decays during fine-tuning.
  • The Robustness Trade-Off: Models trained from scratch on simulation perform exceptionally well on that specific simulation but fail when transferred. RINO accepts a minor performance penalty in-domain (~4-5%) to achieve massive, robust performance gains out-of-domain.

Key Results

When evaluated on the target "real data" domain (JETCLASS), RINO-pretrained models significantly outperformed the supervised baseline trained from scratch on simulation across all model sizes (nano, lite, mini, and base).

Table 1: Classification accuracy comparison across model sizes and training strategies on JETCLASS and JETNET datasets. Results report µ ± σ over 10-fold cross-validation with 80 : 20 train-test splits.
  • Massive Accuracy Gains: Absolute accuracy improvements on the target domain ranged from 15% to 26%. For instance, the supervised base model achieved only 62.9% accuracy on JETCLASS, whereas the RINO-Linear variant reached 80.4%.
  • Naturally Separated Embeddings: Visualizations using PCA and t-SNE demonstrate that RINO's pretrained embedding space naturally clusters top quark decays (Tbqq) away from QCD jets, despite never seeing a top quark label during pretraining.
Figure 2: Visualization of learned jet representations from the base model using PCA (left) and t-SNE embedding (right). Jet representations of all five models are shown in Appendix A.3.

Limitations and Future Work

While RINO successfully tackles domain shift, the authors note a few limitations and areas for development:

  1. Fine-Tuning Sensitivity: Linear probing alone fails to fully exploit the backbone's capacity, while aggressive fine-tuning can overfit and erase the benefits of pretraining. More sophisticated adaptation methods are needed.
  2. Proxy Data: This study used simulated datasets (JETCLASS and JETNET) to emulate domain shift. Future work must validate RINO on actual experimental collision data from the LHC.
  3. Robustness Metrics: More rigorous, diverse metrics are required to fully quantify how robust these self-supervised representations are against various types of systematic detector uncertainties.

Summary updated

Abstract

A common challenge with supervised machine learning (ML) in high energy physics (HEP) is the reliance on simulations for labeled data, which can often mismodel the underlying collision or detector response. To help mitigate this problem of domain shift, we propose RINO (Renormalization Group Invariance with No Labels), a self-supervised learning approach that can instead pretrain models directly on collision data, learning embeddings invariant to renormalization group flow scales. In this work, we pretrain a transformer-based model on jets originating from quantum chromodynamic (QCD) interactions from the JetClass dataset, emulating real QCD-dominated experimental data, and then finetune on the JetNet dataset -- emulating simulations -- for the task of identifying jets originating from top quark decays. RINO demonstrates improved generalization from the JetNet training data to JetClass data compared to supervised training on JetNet from scratch, demonstrating the potential for RINO pretraining on real collision data followed by fine-tuning on small, high-quality MC datasets, to improve the robustness of ML models in HEP.
BibTeX
@misc{hao2025rino,
  author = {Zichun Hao and Raghav Kansal and Abhijith Gandrakota and Chang Sun and Ngadiuba Jennifer and Javier Duarte and Maria Spiropulu},
  title = {{RINO: Renormalization Group Invariance with No Labels}},
  year = {2025},
  eprint = {2509.07486},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.07486},
}
Figure