Calibrated uncertainty estimates for machine-learning surrogates in scientific simulation

Abstract

Machine-learning surrogates are increasingly used to accelerate expensive physical simulations, but their predictive uncertainty is frequently miscalibrated. We introduce a conformal ensemble procedure that provides finite-sample coverage guarantees without retraining the base model. Across five benchmark simulation tasks, the method achieves nominal coverage with intervals up to 35% narrower than existing baselines.

Keywords

• uncertainty quantification
• conformal prediction
• surrogate models