Bayesian Machine Learning Approach to the Quantification of Uncertainties on Ab Initio Potential Energy Surfaces.

Area of Science:

• Computational chemistry
• Quantum mechanics
• Chemical physics

Background:

• Accurate potential energy surfaces (PESs) are crucial for predicting chemical reaction dynamics.
• Quantifying uncertainties in ab initio calculations and their impact on macroscopic properties remains a challenge.

Purpose of the Study:

• To develop and apply a novel methodology for quantifying uncertainties in PESs derived from first-principles calculations.
• To assess the impact of PES uncertainties on microscopic and macroscopic quantities of interest (QoIs).

Main Methods:

• Bayesian inference and machine learning to build stochastic PESs.
• Forward propagation of uncertainties using quasi-classical trajectory and master equation calculations.
• Correlation analysis to identify regions of the PES requiring refinement.

Main Results:

• The methodology successfully quantifies PES uncertainty contributions to QoIs.
• Negligible uncertainty was found for the singlet (1¹A') O2-O PES, with results agreeing with literature.
• Significant uncertainty was identified for the quintet (2⁵A') O2-O PES, particularly concerning the exchange barrier and repulsive wall.

Conclusions:

• The developed methodology provides explicit determination of PES uncertainty impact on chemical kinetics and relaxation properties.
• Uncertainties in the quintet PES lead to a 10-fold variation in vibrational relaxation time and a 4-fold variation in exchange reaction rate coefficient at 2500 K.
• The study highlights the importance of accurate PES description for reliable predictions in combustion and atmospheric chemistry.