Managing Expectations and Imbalanced Training Data in Reactive Force Field Development: An Application to Water...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• ReaxFF is a computationally efficient model for reactive molecular dynamics simulations.
• Optimizing ReaxFF parameters requires careful weighting of diverse training data with varying abundance.
• Current methods face challenges in assigning weights for a balanced compromise among training requirements.

Purpose of the Study:

• To introduce a novel loss function, Balanced Loss, and a workflow for ReaxFF parameter optimization.
• To replace manual weight assignment with a more manageable procedure based on data categories and tolerances.
• To provide a method that validates expectations during optimization and allows for tolerance reconfiguration.

Main Methods:

• Developed a new loss function: Balanced Loss, utilizing a Log-Sum-Exp form.
• Implemented a workflow dividing training data into categories with defined tolerances (acceptable root-mean-square errors).
• Applied the methodology to parametrize ReaxFF for water adsorption on alumina.

Main Results:

• The Balanced Loss methodology successfully parametrized ReaxFF for water adsorption on alumina.
• The resulting force field accurately reproduced both frequent and rare properties from a validation set.
• Demonstrated the robustness of the new force field through molecular dynamics simulations of water desorption.

Conclusions:

• The Balanced Loss approach offers a more manageable and effective procedure for ReaxFF force field optimization.
• This new methodology leads to improved accuracy and robustness in molecular dynamics simulations.
• The developed ReaxFF force field shows promise for studying water-alumina interactions.