Scalable simulation-based inference framework for large-scale validation in fusion

Paper in proceeding, 2025

Simulation-based inference (SBI) is a rapidly developing field aiming to address the inference challenges that are typically related to complex, high-fidelity simulators in science [1]. Computationally costly models, with a selection of uncertain input parameters, are ubiquitous in science, including magnetic confinement fusion research (MCF) [2 – 5]. Due to the input uncertainties, multiple forward passes are typically required in model validation to quantify the input parameter distributions that best reproduce the experimental observations [6]. Solving this inverse problem is one of the key challenges in model validation and must be done algorithmically, such as using SBI, in large-scale validation workflows. Data-efficient, inverse inference workflows have been demonstrated in model validation applications within MCF [6 – 9]. However, a large-scale adoption of these approaches in model validation activities in MCF has not emerged yet. One of the key entry barriers is the software infrastructure that is needed for large-scale applications. In addition to the SBI algorithms, a large-scale validation workflow on high-performance computing (HPC) platforms requires solutions for overall distributed task orchestration, management of the simulation database, and for processing failed simulations without human-in-the-loop. These requirements align with those needed for data generation for general machine learning surrogate modelling of computationally expensive models [10 – 13]. Therefore, in this work, steps are taken to build a scalable SBI framework on top of the Enchanted-surrogates, originally designed for simulation data generation for surrogate models [13].