Machine-learning based prediction of small molecule-surface interaction potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Predicting molecule-surface adsorption affinity is vital across fields like catalysis, drug delivery, and safety.
• Accurate computational prediction is challenging due to complex medium effects.
• Existing methods often lack flexibility for diverse chemical-surface interactions.

Purpose of the Study:

• To develop a flexible machine learning approach for predicting potentials of mean force (PMFs) and adsorption energies.
• To accurately model chemical-surface interactions, considering the influence of the surrounding environment.
• To provide a computationally efficient tool for assessing adsorption affinity.

Main Methods:

• Utilized a machine learning model trained on a pre-existing library of PMFs.
• Employed atomistic molecular dynamics simulations to generate initial PMF data.
• Input features derived from separate interaction potentials of molecules and surfaces with probe atoms.

Main Results:

• Achieved good agreement between predicted and original PMFs on both training and validation datasets.
• Demonstrated the model's predictive power and robustness.
• Successfully generated PMFs for novel molecules and surfaces not included in the training set.

Conclusions:

• The developed machine learning approach offers a flexible and accurate method for predicting molecule-surface adsorption.
• This computational tool can significantly aid research in catalysis, drug delivery, and materials design.
• The model's ability to generalize to unseen data highlights its practical utility.