Cut-SOAP: A Machine Learning Descriptor for Rapid Screening of Molecular Adsorption Energetics

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate adsorption energy prediction is vital for catalysis and chemical reaction studies.
• Conventional quantum chemistry methods are computationally expensive for large-scale screening.
• There is a need for efficient and accurate methods to predict adsorption energies.

Purpose of the Study:

• To develop a machine learning (ML) pipeline for efficient prediction of relative energy interactions in molecular adsorption.
• To reduce the computational cost associated with traditional quantum chemistry methods.
• To enable rapid screening and accelerate the discovery of catalytic systems.

Main Methods:

• Utilized a modified Smooth Overlap of Atomic Positions (SOAP) descriptor, termed Cut-SOAP, to reduce feature dimensionality by over 97%.
• Constructed a large adsorption dataset of over 430,000 entries using Fritz-Haber Institute ab initio materials simulation (FHI-aims) output data.
• Trained a deep neural network on the dataset, optimizing architecture and hyperparameters for accuracy and computational efficiency.

Main Results:

• Achieved a mean absolute error below 0.1 eV on the standard test set.
• Maintained a robust mean absolute error below 1.0 eV on a challenging out-of-distribution dataset, demonstrating strong generalization.
• The trained model can perform thousands of predictions in seconds, highlighting its efficiency for rapid screening.

Conclusions:

• The proposed ML pipeline offers an accessible, efficient, and accurate tool for predicting relative energy interactions.
• This approach significantly lowers computational costs, making it practical for screening large datasets and complex systems.
• The ML pipeline is a crucial step towards accelerated discovery and optimization of catalytic materials.