Outlier-detection for reactive machine learned potential energy surfaces.

Area of Science:

• Computational chemistry
• Chemical physics
• Machine learning applications in science

Background:

• Accurate potential energy surfaces (PESs) are crucial for simulating chemical reactions.
• Identifying unreliable predictions in PESs is essential for efficient model refinement.
• Uncertainty quantification (UQ) offers a framework for assessing prediction confidence.

Purpose of the Study:

• To evaluate different UQ methods for detecting outliers in reactive molecular PESs.
• To compare the performance of Ensembles, deep evidential regression (DER), and Gaussian Mixture Models (GMM) for UQ.
• To identify strategies for improving the accuracy and efficiency of PES development.

Main Methods:

• Application of three UQ methods: Ensembles, DER, and GMM.
• Testing on the H-transfer reaction between syn-Criegee and vinyl hydroxyperoxide.
• Analysis of outlier detection performance based on prediction uncertainty.

Main Results:

• Ensemble models demonstrated the highest accuracy in detecting outliers (~90% for top 25 structures).
• GMM showed moderate performance (~50% for top 1000 structures).
• DER's performance was significantly limited by its statistical assumptions.
• A structure-based indicator correlated with average error was identified.

Conclusions:

• Ensemble-based UQ is a robust approach for identifying unreliable data points in PES calculations.
• GMM offers a viable alternative for UQ in this context.
• The identified structure-based indicator can accelerate the refinement of neural network models for PESs.