Enhanced Encoding Module of AisNet with Charge Features: An Application to the S_N2 Data Set.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate prediction of molecular forces is crucial for simulating chemical reactions and material properties.
• Existing machine learning models often struggle with capturing long-range interactions and achieving high accuracy across diverse chemical environments.
• Multifeature fusion frameworks offer potential for improved performance by integrating various atomic descriptors.

Purpose of the Study:

• To develop and evaluate a novel charge-enhanced encoding module for the AisNet framework.
• To improve the force error accuracy and cross-model transferability of machine learning force fields.
• To investigate the impact of Coulombic interactions on the resolution of atomic environments.

Main Methods:

• Integration of a Coulomb matrix into the AisNet framework to create AisNet-C.
• Utilizing t-SNE and PCA for analyzing the impact of Coulombic interactions on feature resolution.
• Testing the enhanced module with various model architectures, including PAINN and PhysNet, on the S_N2 dataset.

Main Results:

• AisNet-C demonstrated superior force error accuracy and cross-model transferability compared to the original AisNet.
• The Coulomb matrix effectively enhanced global environmental resolution, synergizing with other features.
• The PAINN model with the enhanced module achieved state-of-the-art performance (0.00423 eV/Å), reducing mean error by 41.8% compared to PhysNet.

Conclusions:

• The charge-enhanced encoding module is a valuable plug-and-play component for improving machine learning force fields.
• Coulombic interactions play a significant role in enhancing the resolution of atomic environments in multifeature fusion models.
• This approach offers a promising direction for developing physically informed machine learning force fields with broader applicability.