Constructing Grids for Molecular Quantum Dynamics Using an Autoencoder.

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning

Background:

• Molecular quantum dynamics (QD) calculations face the curse of dimensionality, limiting their application to complex systems.
• Reducing dimensionality to relevant coordinates is crucial but challenging due to unintuitive coordinate identification.

Purpose of the Study:

• To develop a machine learning approach for identifying relevant coordinates in molecular systems.
• To enable efficient quantum dynamics calculations on reduced-dimensional representations.

Main Methods:

• Utilized an autoencoder trained on molecular configurations from trajectory calculations.
• Generated a low-dimensional potential energy surface grid within the identified subspace.
• Employed the G-matrix formalism for kinetic energy operator calculation on the grid.

Main Results:

• Successfully identified a low-dimensional representation for molecular configurations.
• Demonstrated the feasibility of performing quantum dynamics calculations on the generated grid.
• Provided a practical methodology for grid construction and application.

Conclusions:

• The presented machine learning approach effectively addresses the curse of dimensionality in QD.
• This method facilitates more efficient and accurate quantum dynamics simulations for reactive systems.