Uncertainty quantification for predictions of atomistic neural networks

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Neural networks (NNs) are increasingly used for quantum chemical predictions.
• Quantifying uncertainty in NN predictions is crucial for reliable results.
• The PhysNet architecture was adapted for uncertainty quantification.

Purpose of the Study:

• To quantitatively explore uncertainty quantification (UQ) for trained NNs on quantum chemical data.
• To evaluate a modified PhysNet model (PhysNet-DER) for calibration, prediction quality, and error-uncertainty correlation.
• To investigate the impact of training data composition on NN performance.

Main Methods:

• Modification of the PhysNet NN architecture to PhysNet-DER.
• Evaluation using various metrics for calibration and prediction quality.
• Analysis of error-uncertainty relationships and the influence of training data distribution and chemical space coverage.
• Application to tautomerization reactions and analysis of feature space distances.

Main Results:

• Error and uncertainty were not linearly related, but variance offered insights into training data quality.
• Noisy and redundant training data complicate molecular property predictions, even for minor structural changes.
• Training data composition critically affects prediction accuracy, with implications for chemical databases.
• Specific examples showed how data redundancy or similarity impacts error and variance.

Conclusions:

• UQ provides valuable insights into the reliability of NN predictions in quantum chemistry.
• The composition and quality of training data are paramount for accurate molecular property prediction using NNs.
• The developed method enables information-based database improvement via active learning for targeted applications.