Fast and Accurate Uncertainty Estimation in Chemical Machine Learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning models are increasingly used for predicting atomic and molecular properties.
• Quantifying the uncertainty in these predictions is crucial for reliable applications.
• Existing methods for uncertainty estimation can be computationally expensive or less accurate.

Purpose of the Study:

• To present an inexpensive and reliable scheme for estimating prediction uncertainty in machine learning models.
• To demonstrate the applicability and effectiveness of the proposed method across various chemical and material property predictions.

Main Methods:

• The proposed scheme relies on data resampling, generating multiple models from subsampled training data.
• Maximum likelihood estimation is used to benchmark uncertainty prediction accuracy and correct for model correlations.
• Cross-validation is employed to further enhance the performance of uncertainty estimation.
• The method is particularly efficient for sparse Gaussian Process Regression models.

Main Results:

• The resampled uncertainty estimator is evaluated at negligible cost for sparse Gaussian Process Regression.
• Reliable uncertainty estimates were demonstrated for molecular and materials energetics predictions.
• The method successfully estimated uncertainties in nuclear chemical shieldings for molecular crystals.
• Extension to correlated predictions like energy differences and forces is straightforward.

Conclusions:

• The presented resampling scheme offers an inexpensive and reliable way to estimate uncertainty in machine learning predictions for chemical and material properties.
• This method can be readily applied to diverse machine-learning approaches, improving data-driven predictions.
• The technique facilitates training-set optimization and active-learning strategies, advancing AI in scientific discovery.