A practical guide to simulation-based inference for science
A new tutorial preprint provides researchers with practical tools to apply Simulation-Based Inference (SBI) - machine learning methods that connect complex simulators with data, enabling faster and more robust parameter estimation in domains from astrophysics to neuroscience.
When models are too complex for standard statistical tools, scientists often struggle to connect simulations with experimental data. A new tutorial preprint, “Simulation-Based Inference: A Practical Guide”, provides researchers with a structured entry point into Simulation-Based Inference (SBI): a family of methods that uses machine learning to make parameter estimation feasible even for highly complex simulators.
Instead of requiring explicit likelihoods (i.e. a statistical description on how the data would look like given the simulation parameters), SBI trains neural networks directly on simulated data. Once trained, these networks can be reused to quickly perform Bayesian inference on new observations, making SBI attractive in fields where generating simulations is easier than computing probabilities.
The authors outline a step-by-step workflow, from defining priors and simulators to training inference networks and validating results. They illustrate the process with case studies in astrophysics, psychophysics, and neuroscience, highlighting how SBI can accelerate discovery across scientific domains.
The tutorial represents a broad international collaboration, bringing together researchers from the University of Tübingen and its Tübingen AI Center, Helmholtz AI and Helmholtz-Zentrum Dresden-Rossendorf, several European universities and research institutes including Université Paris-Saclay, Univ. Grenoble Alpes, University of Amsterdam, KU Leuven, VIB-Neuroelectronics Research Flanders, as well as Google Research and the Max Planck Institute for Intelligent Systems.
The guide aims to help domain researchers adopt SBI methods in practice, offering not only algorithms, but also diagnostic tools and best practices to avoid common pitfalls.
Peter Steinbach, head of the Helmholtz AI Consultant Team at Helmholtz-Zentrum Dresden-Rossendorf and co-author, shares his perspective on the SBI tutorial and its relevance for scientists looking to apply these methods in practice:
“High fidelity simulators are everywhere - especially in physics. Due to its first-principle based way of approaching scientific insight, the physics community has access to a plethora of in-silico simulators which help to predict how data would look in experiments. This tutorial paper on SBI provides a direct and hands-on introduction on how to use these modern AI driven tools. SBI can help lift experimental inference to unprecedented levels of precision and trustworthiness in scientific domains where simulators are at hand. I am glad that the wonderful SBI community made this paper happen.”
The tutorial does not aim to reinvent statistical inference but to make established SBI methods easier to use in practice. By offering workflows, diagnostics, and concrete examples, the authors hope this guide will help scientists who rely on simulations explore whether SBI can support their own research questions.