Speaker
Description
Training state-of-the-art neural networks for high-energy physics (HEP) tasks typically requires massive, fully simulated datasets—whose generation is both computationally expensive and experiment-specific. In this work, we demonstrate that this dependence on large-scale full simulations can be drastically reduced by leveraging pretrained models trained on fast-simulation data. These pretrained models serve as foundation models that can be fine-tuned for diverse downstream tasks, even across different experimental setups.
We validate this approach using a set of representative HEP tasks—event classification, jet tagging, and missing transverse energy reconstruction—spanning a variety of modern network architectures. Our results show that models initialized from fast-simulation pretraining achieve competitive or superior performance with only a small fraction of full-simulation data, reducing the need for costly event generation by up to an order of magnitude. Moreover, we observe consistent transferability across experiments, highlighting the potential for cross-collaboration reuse of pretrained HEP models.
This study suggests a path toward data-efficient and simulation-aware machine learning in particle physics, paving the way for scalable AI pipelines that are both sustainable and experiment-agnostic.