This paper proposes a comprehensive uncertainty quantification framework for emulated geohazard runout predictions, integrating Gaussian Process emulation, Deep Ensembles, Monte Carlo Dropout, and Conformal Prediction to characterize aleatoric, epistemic uncertainties, and emulator approximation error.
Key findings
A unified taxonomy of uncertainty sources in emulated geohazard predictions and corresponding quantification strategies.
Integration of complementary UQ methods with guidance on method selection based on application requirements.
Novel extensions for spatial uncertainty quantification, including spatially varying length-scales in GP kernels and conformalized spatial predictions.
Comprehensive validation protocol using r.avaflow simulations, including systematic comparison metrics and ablation studies.
Limitations & open questions
Limited training data for surrogate models.
High-dimensional parameter spaces and spatially distributed outputs pose challenges for UQ methods.
Distribution-free guarantees are needed for reliable confidence intervals in complex geohazard simulators.