Uncertainty Prediction for Machine Learning Models of Material Properties
This work addresses the need for reliable uncertainty quantification in AI applications for material science, though it is incremental as it compares existing approaches.
The paper compared three methods for predicting individual uncertainty intervals in machine learning models of material properties, finding that directly modeling uncertainties was easiest to fit and minimized error over- and under-estimation in most cases across 12 properties.
Uncertainty quantification in Artificial Intelligence (AI)-based predictions of material properties is of immense importance for the success and reliability of AI applications in material science. While confidence intervals are commonly reported for machine learning (ML) models, prediction intervals, i.e., the evaluation of the uncertainty on each prediction, are seldomly available. In this work we compare 3 different approaches to obtain such individual uncertainty, testing them on 12 ML-physical properties. Specifically, we investigated using the Quantile loss function, machine learning the prediction intervals directly and using Gaussian Processes. We identify each approachs advantages and disadvantages and end up slightly favoring the modeling of the individual uncertainties directly, as it is the easiest to fit and, in most cases, minimizes over-and under-estimation of the predicted errors. All data for training and testing were taken from the publicly available JARVIS-DFT database, and the codes developed for computing the prediction intervals are available through JARVIS-Tools.