With the increase in luminosity and detector granularity, simulation will be a significant computational challenge in the upcoming high-luminosity era of the CERN Large Hadron Collider (LHC). To tackle this, I present developments in graph- [1, 2] and attention-based [3, 4] machine learning (ML) models for generating jets at the LHC using sparse and efficient point cloud representations of our data, which offer a three-orders-of-magnitude improvement in latency compared to full simulations. I also present studies on metrics for validating ML-based simulations, including the novel Fréchet and kernel physics distances, which are found to be highly sensitive to typical mismodelling by ML generative models [3]. Finally, I discuss directions for future work, including incorporation of symmetries such as Lorentz-equivariance [5].
[1] RK et al, ML4PS @ NeurIPS 2020, https://arxiv.org/abs/2012.00173
[2] RK et al, NeurIPS 2021, https://arxiv.org/abs/2106.11535
[3] RK et al, PRD 2023, https://arxiv.org/abs/2211.10295
[4] A. Li, V. Krishnamohan, RK et al, ML4PS @ NeurIPS 2023, https://arxiv.org/abs/2312.04757
[5] Z. Hao, RK et al, EPJC 2023, https://arxiv.org/abs/2212.07347