Erlangen 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 50: Data, AI, Computing, Electronics V

T 50.1: Talk

Wednesday, March 18, 2026, 16:15–16:30, KH 00.024

NEEDLE - A modern orchestration framework for Neural Simulation Based Inference tools — •Kylian Schmidt¹, Ulrich Husemann², Nicoló Trevisani¹, Stephen Jiggins², Levi Evans², Judith Katzy², Stefan Katsarov², Felix Kahlhöfer³, Nino Kovačić⁴, Lena Rathmann³, and Niklas Reus³ — ¹ETP, KIT, Karlsruhe — ²DESY, Hamburg — ³IAP, KIT, Karlsruhe — ⁴Department of Physics, U. of Zagreb

Neural Simulation Based Inference (NSBI) is a new machine learning (ML) paradigm for statistical data analysis in high energy physics (HEP). These tools learn the underlying statistical distribution of the data using surrogate neural networks and show clear improvements over classical likelihood estimation methods. However, these methods require the training of a large number of networks in order to achieve this increase in performance while retaining robustness towards biases. It is therefore crucial to address the challenges of orchestrating many neural network trainings, alongside efficient utilization of computational resources.

The NEEDLE project aims to provide a flexible framework for distributed training together with a library of NSBI tools for deployment on High Performance Computing clusters. NEEDLE combines modern ML libraries together with commonly-used HEP data analysis tools and formats. In this talk, we present the design of the NEEDLE framework, how it handles large data streams and our integration with pytorch lightning.

Keywords: Simulation Based Inference; Machine Learning; High Performance Computing; Workflow Management; Neural Networks