3D Spatial Learning for Adsorption Energy Prediction in Multi-Temporal Solution Systems: The MTSS Data Set and a...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Current adsorption energy prediction methods struggle with dynamic solution systems and diverse spatial configurations.
• Traditional datasets are static, failing to capture the conformational space explored by dynamic systems over time.

Purpose of the Study:

• To introduce the Multi-Temporal Solution System (MTSS) dataset for temporally resolved adsorption prediction.
• To develop a novel deep learning model, SEP-Net, capable of modeling solution-level interactions.

Main Methods:

• Creation of the MTSS dataset with 500,000 temporally resolved configurations and adsorption energy labels across five solvents.
• Proposal of SEP-Net, a dual-channel graph network integrating rotational-invariant geometric learning and molecular SMILES embeddings.
• Experimental validation of SEP-Net's performance against traditional methods like MLP.

Main Results:

• SEP-Net achieved a Mean Absolute Error (MAE) of 211.02 kJ/mol on known solvents and 507.37 kJ/mol on unseen solvents.
• SEP-Net significantly outperformed MLP, demonstrating a substantial improvement in prediction accuracy (e.g., 3827.33 vs 507.37 kJ/mol on ACE solvent).

Conclusions:

• The MTSS dataset and SEP-Net establish new benchmarks for system-level adsorption prediction.
• Geometric deep learning effectively addresses the complexities of solute-solvent and solvent-solvent interactions in dynamic systems.