Using principal component analysis for neural network high-dimensional potential energy surface.

Area of Science:

• Computational Chemistry
• Chemical Physics
• Machine Learning in Chemistry

Background:

• Potential energy surfaces (PESs) are crucial for understanding chemical reactions.
• Calculating accurate PESs is computationally demanding for complex systems.
• Artificial neural networks (NNs) show promise for PES construction, but require effective molecular descriptors.

Purpose of the Study:

• To develop an optimized protocol for constructing accurate and efficient NNs for PES calculations.
• To address the bottleneck of selecting suitable molecular descriptors for NNs.
• To improve the learning and predictive capabilities of NNs for chemical reaction dynamics.

Main Methods:

• Utilized principal component analysis (PCA) to identify an optimal set of molecular descriptors.
• Employed PCA to reduce the dimensionality of the input space for NNs without sacrificing accuracy.
• Applied the developed protocol to model the high-dimensional PES for the keto-enol tautomerism of acetone.

Main Results:

• PCA effectively prepared an optimal set of descriptors for NN-based PES construction.
• The PCA-enhanced NN protocol demonstrated substantial improvements in learning and prediction accuracy.
• Reduced input space dimensionality via PCA led to more efficient NN models.

Conclusions:

• PCA is a powerful tool for optimizing molecular descriptors in NN-based PES calculations.
• This novel approach enhances the efficiency and accuracy of predicting chemical reaction dynamics.
• The method shows promise for applications in complex chemical systems, such as tautomerism reactions.