Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interdisciplinary Digital Twin Engine InterTwin for calorimeter simulation

Corentin Allaire, Vera Maiboroda, David Rousseau

TL;DR

Geant4 calorimeter simulations are computationally expensive, limiting statistical reach for precision measurements. The approach uses CaloINN, an invertible neural network based on normalizing flows, within the interTwin Digital Twin Engine and pursues tail-focused corrections using tail-enhanced training and density-ratio reweighting to match the Geant4 tails; the conditional distribution to model is $p(I|E_{inc})$. The study demonstrates tail-targeted accuracy goals of better than 10% in the tails at the $10^{-3}$ quantile while preserving speed-ups of 2–3 orders of magnitude. This production-ready, open-source digital twin framework for accelerated calorimeter simulation has potential to scale to large HL-LHC workloads and cross-domain DTE applications.

Abstract

Calorimeter shower simulations are computationally expensive, and generative models offer an efficient alternative. However, achieving a balance between accuracy and speed remains a challenge, with distribution tail modeling being a key limitation. Invertible generative network CaloINN provides a trade-off between simulation quality and efficiency. The ongoing study targets introducing a set of post-processing modifications of analysis-level observables aimed at improving the accuracy of distribution tails. As part of interTwin project initiative developing an open-source Digital Twin Engine, we implemented the CaloINN within the interTwin AI framework.

Interdisciplinary Digital Twin Engine InterTwin for calorimeter simulation

TL;DR

Geant4 calorimeter simulations are computationally expensive, limiting statistical reach for precision measurements. The approach uses CaloINN, an invertible neural network based on normalizing flows, within the interTwin Digital Twin Engine and pursues tail-focused corrections using tail-enhanced training and density-ratio reweighting to match the Geant4 tails; the conditional distribution to model is . The study demonstrates tail-targeted accuracy goals of better than 10% in the tails at the quantile while preserving speed-ups of 2–3 orders of magnitude. This production-ready, open-source digital twin framework for accelerated calorimeter simulation has potential to scale to large HL-LHC workloads and cross-domain DTE applications.

Abstract

Calorimeter shower simulations are computationally expensive, and generative models offer an efficient alternative. However, achieving a balance between accuracy and speed remains a challenge, with distribution tail modeling being a key limitation. Invertible generative network CaloINN provides a trade-off between simulation quality and efficiency. The ongoing study targets introducing a set of post-processing modifications of analysis-level observables aimed at improving the accuracy of distribution tails. As part of interTwin project initiative developing an open-source Digital Twin Engine, we implemented the CaloINN within the interTwin AI framework.

Paper Structure

This paper contains 6 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Choosing a model for fast calorimeter simulation
  3. interTwin project
  4. Fixing distribution tails
  5. Conclusion
  6. Funding information

Figures (1)

  • Figure 1: Distribution of voxel energy compared between Geant4 (black) and CaloINN (blue) on the CaloChallenge dataset containing pions. While the bulks of distributions are in a good agreement, some discrepancy can be observed in the tails.