A dual-cutoff machine-learned potential for condensed organic systems obtained via uncertainty-guided active learning

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine-learned potentials (MLPs) bridge the gap between classical and first-principles methods.
• Accurate modeling of organic compounds in condensed phases remains a challenge.

Purpose of the Study:

• To develop and train an MLP for accurate potential energy surface and property prediction of organic molecules.
• To create a versatile descriptor for both intramolecular and intermolecular interactions.

Main Methods:

• Implementation and training of a novel dual descriptor based on the atomic cluster expansion (ACE).
• Uncertainty-guided active learning for efficient training set generation.
• Application to alcohols, alkanes, and adipates in molecular and condensed phases.

Main Results:

• MLP accurately predicts densities with <4% discrepancy to experiment.
• Vibrational frequencies show <1 THz RMSE compared to DFT.
• Heat capacities of condensed systems are within 11% of experimental values.

Conclusions:

• The dual descriptor accurately captures short-range intramolecular and long-range intermolecular interactions.
• MLPs trained with this approach provide a computationally efficient and accurate tool for organic materials.
• The method demonstrates broad applicability for diverse organic compounds.