Equivariant Neural Networks Utilizing Molecular Clusters for Accurate Molecular Crystal Lattice Energy Predictions

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Equivariant neural networks offer high data efficiency and generalization for interatomic potentials.
• Traditional density functional theory (DFT) methods for crystal structure prediction are computationally expensive.
• Predicting organic molecular crystal structures requires accurate and efficient methods.

Purpose of the Study:

• To adapt equivariant neural networks for predicting molecular crystal structure energies.
• To reduce the computational cost associated with crystal structure property calculations.
• To enhance the discovery of low-energy crystal structures.

Main Methods:

• Utilized the Allegro equivariant neural network architecture.
• Trained the network on molecular clusters using a Gaussian-type orbital (GTO)-based method.
• Integrated the trained Allegro network with the USPEX crystal structure prediction framework.

Main Results:

• Achieved accurate prediction of molecular crystal structure energies at reduced computational cost.
• Demonstrated that the network's predictions align closely with plane-wave DFT results.
• Successfully accelerated the discovery of low-energy crystal structures through framework integration.

Conclusions:

• Equivariant neural networks provide a computationally efficient and accurate alternative for predicting molecular crystal properties.
• This approach bypasses the need for resource-intensive periodic DFT calculations.
• The integration with USPEX significantly speeds up the crystal structure prediction workflow.