Neural Bayesian Online Calibration with Deficient Simulators for MADMAX Dark Matter Experiment
Scientific experiments often rely on deficient simulators and hardware whose commanded settings differ from the states realized in practice, limiting reliable optimization and inference. A Bayesian online calibration framework is introduced for experiments with at least one directly measurable simulator output and potentially latent quantities of scientific interest. Pre-calibration data are used to learn state-dependent effective simulator parameters, while a normalizing flow enables amortized posterior inference from measurement histories. During calibration, measurement and positioning uncertainties are propagated, and hardware commands are selected by balancing expected improvement with information gain. The framework is applied to the MADMAX dark matter experiment, where measured reflectivity is used to calibrate a deficient disk-booster simulator and infer the unobservable axion-signal power boost factor. Experimental validation on the MADMAX experiment is ongoing.