Reliability of Probabilistic Emulation of Physical Systems

Name: Reliability of Probabilistic Emulation of Physical Systems
Uploaded: 2026-10-06T01:51:31.235950+00:00
Duration: 442 s
Description: The study compared two methods for forecasting physical systems: generative models and ensembles of deterministic models trained with CRPS loss. The findings show that CRPS-trained ensembles provide more reliable uncertainty estimates across single-step predictions and long-term rollouts, outperforming generative models in coverage.

The study compared two methods for forecasting physical systems: generative models and ensembles of deterministic models trained with CRPS loss. The findings show that CRPS-trained ensembles provide more reliable uncertainty estimates across single-step predictions and long-term rollouts, outperforming generative models in coverage.

Open the full episode