Simulation-Based Inference: Robust Methods for Astronomy and the Broader Sciences

1630 

Topic-Contributed Paper Session 

Maximilian Autenrieth (1), Maximilian Autenrieth (1), Konstantin Karchev (2), Kaisey Mandel (3), Lishan Shi (4), Matthew O'Callaghan (3), James Carzon (1)

(1) N/A, N/A, (2) International School for Advanced Studies (SISSA) & University of Barcelona, N/A, (3) University of Cambridge, N/A, (4) The Pennsylvania State University, N/A

Maximilian Autenrieth  
N/A

Maximilian Autenrieth  
N/A

Konstantin Karchev  
International School for Advanced Studies (SISSA) & University of Barcelona

Kaisey Mandel  
University of Cambridge

Lishan Shi  
The Pennsylvania State University

Matthew O'Callaghan  
University of Cambridge

James Carzon  
N/A

Simulators and generative models have become indispensable research tools across science and engineering, ranging from mechanistic descriptions of physical systems to flexible data driven representations. Enabled by high performance computing and advances in machine learning, Simulation-Based Inference (SBI) has emerged as a powerful direction for drawing statistical conclusions in complex inverse problems where likelihood based methods are intractable or prohibitively expensive.

Approximate Bayesian Computation has long provided a principled framework for inference without tractable likelihoods. Recent SBI methods complement these ideas by incorporating flexible neural density and ratio estimators that support scalable and amortized inference with models of high realism and high dimensional data. Although astronomy has been a major driver of practical uptake and methodological innovation, SBI has been been advanced across the sciences by challenges arising in ecology, genetics, epidemiology and engineering, where complex simulators produce data whose likelihoods are analytically inaccessible (e.g., see https://simulation-based-inference.org/). These challenges make SBI a timely topic for the broader statistics community.

This session highlights statistical themes that arise in SBI, including calibration and diagnostics for neural posteriors, identifiability in implicit models, uncertainty quantification under model misspecification, and connections between classical and simulation-based inference. Aiming to demonstrate SBI's potential with concrete scientific examples and to point out areas where statistical insight is essential, the session comprises five speakers with nominal titles:

1. An Introduction to SBI for Astronomical Applications (Kaisey Mandel, Cambridge)
2. Fully Simulation-Based Transient Science with LSST (Konstantin Karchev, SISSA & University of Barcelona)
3. Millions of Galaxies, Sparse Information: Reliable SED Inference for HETDEX with Neural Density Estimators (Lishan Shi, Penn State)
4. Robust and Scalable Simulation-Based Inference for Stellar Parameters via Domain Adaptation (Matthew O'Callaghan, Cambridge)
5. Trustworthy Scientific Inference from Limited or Sparse Calibration Data (James Carzon, Carnegie Mellon)


Progress in SBI reflects collaboration between statisticians, domain scientists and machine learning researchers. Modern SBI methods offer scalability and amortization and can accommodate realistic data generating processes that are difficult to represent within classical likelihood based analysis, for example when selection effects or complex measurement processes are present. As a result SBI is becoming a practical tool in many areas of applied science. Strengthening its methodological foundations is important for statistical practice, ensuring robustness and reliability in the spirit of classical inference.

The session appeals to a broad audience across the sciences and applications, engaging both Bayesian and frequentist perspectives as well as the statistical learning community. By enabling principled inference from complex simulations, SBI opens new directions for statistical theory, methodology and practice.

Astrostatistics Interest Group ³

Section on Bayesian Statistical Science ²

Section on Physical and Engineering Sciences ¹