Abstract

As safety-critical systems—ranging from autonomous transport to industrial control systems—become increasingly data-driven, ensuring reliable probabilistic predictions is a fundamental challenge. Historically, the safety of such systems has been ensured through physics-based modeling, scenario analysis, and conservative engineering design. However, as machine learning (ML) models are increasingly used for predictive decision-making, they introduce additional uncertainties that must be well-calibrated to maintain system reliability. This talk explores the role of probabilistic calibration techniques in improving the trustworthiness of ML-based predictions in safety-critical applications. We will discuss:

The challenges of uncalibrated ML models in high-risk environments.

Approaches for calibrating ML predictions, from conformal prediction to Bayesian calibration.

The role of uncertainty-aware experimental design to reduce uncertainty in safety-critical applications.

Hybrid approaches that combine physics-based models with data-driven insights to ensure robustness.