Trade-offs in field-level simulation-based inference

Characterize the trade-offs of field-level simulation-based inference in astronomy by determining under what conditions using full field-level information with neural network-based density estimators provides genuine constraint improvements while avoiding overfitting to simulation-specific features, and by establishing validation protocols that ensure reliable generalization beyond the training simulations.

Background

Simulation-based inference enables Bayesian analysis when analytical likelihoods are intractable, including field-level approaches that use entire spatial fields rather than summary statistics. While these methods can potentially extract non-Gaussian information, neural networks risk learning artifacts specific to the training simulations, undermining generalization.

The review highlights that, for field-level inference, the balance between potential gains and risks is not yet well understood. This uncertainty necessitates precise characterization of when field-level methods outperform summary-statistic approaches and how to validate them, especially under domain shifts across different simulations and observational systematics.

References

For field-level inference specifically, the trade-offs remain unclear.

— Deep Learning in Astrophysics  (2510.10713 - Ting, 12 Oct 2025) in Section 3.2.6, A Cautionary Note: On The Black Box Critique

Moreover, because the selected summaries are not known to be sufficient, the resulting ABC posterior depends on their choice and may differ from the posterior based on the full plume observations.

— Bayesian inversion of multilayer $\mathrm{CO}_2$ migration from seismic plume observations using a graph-based finite-rate invasion-percolation model  (2609.04937 - Svee et al., 4 Sep 2026) in Section 3.3, “Plume-outline summary statistics”