A Simple and Scalable Kernel Density Approach for Reliable Uncertainty Quantification in Atomistic Machine Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) models are vital for predicting material properties and accelerating simulations.
• Model reliability hinges on the representativeness of training data, posing challenges in high-dimensional spaces.
• Current methods for uncertainty quantification can be computationally expensive, often requiring model ensembles.

Purpose of the Study:

• To develop a scalable and efficient uncertainty quantification (UQ) framework for ML models in materials science.
• To provide a model-agnostic metric for assessing prediction trustworthiness.
• To enable practical deployment and improve the interpretability of ML models in scientific applications.

Main Methods:

• Implemented a GPU-accelerated UQ framework utilizing k-nearest-neighbor kernel density estimation (KDE).
• Employed principal component analysis (PCA) for dimensionality reduction in descriptor space.
• Validated the framework across diverse chemical systems, ML models, descriptors, and material properties.

Main Results:

• The KDE-based uncertainty score effectively identifies sparsely sampled regions and extrapolative configurations.
• The UQ metric demonstrates strong correlation with traditional ensemble-based uncertainty measures.
• The framework successfully highlights areas where ML model predictions are less trustworthy.

Conclusions:

• The proposed KDE-based UQ framework offers a practical and transferable solution for assessing ML model reliability.
• It enhances the interpretability and robustness of ML models in materials science.
• This approach facilitates the deployment readiness of ML models by quantifying prediction uncertainty.