Abstract

Johannes Dietschreit

University Vienna

I will discuss uncertainty quantification (UQ) and active learning (AL) for machine learning models, with a focus on their application to machine-learned interatomic potentials (MLIPs).

Active learning is a powerful method for improving machine learning models by selecting the most informative data points for training. By targeting regions of high uncertainty, AL enables more efficient data acquisition, minimizing the amount of data needed to achieve high model performance. This presentation will explore different uncertainty measures and their impact on model accuracy and robustness.

A key application of AL lies in the development of machine-learned interatomic potentials. Here, UQ measures play a key role in guiding the exploration of configurational space. When the gradient of the uncertainty is available, it can be used to actively steer the molecular system toward configurations that the model is least certain about. This strategy accelerates the discovery of new and informative data, ultimately improving the predictive capabilities of MLIPs.