Hybrid classical/machine-learning force fields for the accurate description of molecular condensed-phase systems

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Electronic structure methods provide accurate molecular property predictions but are computationally expensive.
• Classical force fields (FFs) are efficient but rely on approximations and extensive training data, a challenge for condensed-phase systems.
• Machine learning (ML) force fields further amplify data requirements, especially for complex condensed-phase simulations.

Purpose of the Study:

• To develop a hybrid ML/classical FF model that is transferable to condensed-phase systems.
• To overcome the limitations of data generation for training ML FFs in condensed phases.
• To combine the efficiency of classical FFs with the adaptability of ML.

Main Methods:

• Parametrization of a hybrid ML/classical FF model using high-quality *ab initio* data of dimers and monomers in vacuum.
• Integration of ML corrections into a classical FF framework to address limitations of classical approximations.
• Validation of the model on benchmarking datasets and experimental condensed-phase data.

Main Results:

• The proposed hybrid model demonstrates transferability to condensed-phase systems.
• The model effectively combines the robustness of classical FFs with the flexibility of ML.
• Validation on organic liquids and small-molecule crystal structures confirms the model's performance.

Conclusions:

• The developed hybrid ML/classical FF model offers a promising approach for FF development.
• This method can potentially unlock the full capabilities of classical FFs for complex systems.
• The strategy addresses the challenge of data scarcity in ML-driven FF development for condensed-phase simulations.