Considerations in the use of ML interaction potentials for free energy calculations

Orlando A. Mendible,Jonathan K. Whitmer,Yamil J. Colón

2024-03-21

Abstract:Machine learning potentials (MLPs) offer the potential to accurately model the energy and free energy landscapes of molecules with the precision of quantum mechanics and an efficiency similar to classical simulations. This research focuses on using equivariant graph neural networks MLPs due to their proven effectiveness in modeling equilibrium molecular trajectories. A key issue addressed is the capability of MLPs to accurately predict free energies and transition states by considering both the energy and the diversity of molecular configurations. We examined how the distribution of collective variables (CVs) in the training data affects MLP accuracy in determining the free energy surface (FES) of systems, using Metadynamics simulations for butane and alanine dipeptide (ADP). The study involved training forty-three MLPs, half based on classical molecular dynamics data and the rest on ab initio computed energies. The MLPs were trained using different distributions that aim to replicate hypothetical scenarios of sampled CVs obtained if the underlying FES of the system was unknown. Findings for butane revealed that training data coverage of key FES regions ensures model accuracy regardless of CV distribution. However, missing significant FES regions led to correct potential energy predictions but failed free energy reconstruction. For ADP, models trained on classical dynamics data were notably less accurate, while ab initio-based MLPs predicted potential energy well but faltered on free energy predictions. These results emphasize the challenge of assembling an all-encompassing training set for accurate FES prediction and highlight the importance of understanding the FES in preparing training data. The study points out the limitations of MLPs in free energy calculations, stressing the need for comprehensive data that encompasses the system's full FES for effective model training.

Chemical Physics,Materials Science,Machine Learning

What problem does this paper attempt to address?

The paper discusses the application of machine learning potentials (MLPs) in free energy calculations, specifically whether they can accurately reproduce the free energy surfaces (FES) of systems. The study analyzes the impact of the distribution of collective variables (CVs) in the training data on the accuracy of predicting FES, by comparing classical molecular dynamics and MLPs trained from first principles data. The results indicate that if the training data does not sufficiently represent all the feature regions of the FES, the model may not accurately reconstruct the free energy of certain configurations. For more complex systems such as alanine dipeptide (ADP), even though MLPs perform well in potential energy prediction, they still face difficulties in predicting free energy, emphasizing the importance of generating comprehensive training datasets.

Considerations in the use of ML interaction potentials for free energy calculations

Accelerating Metadynamics-Based Free-Energy Calculations with Adaptive Machine Learning Potentials.

Stable and Accurate Atomistic Simulations of Flexible Molecules using Conformationally Generalisable Machine Learned Potentials

Discrepancies and Error Evaluation Metrics for Machine Learning Interatomic Potentials

Challenges for Machine Learning Force Fields in Reproducing Potential Energy Surfaces of Flexible Molecules

Accurate Free Energy Calculation via Multiscale Simulations Driven by Hybrid Machine Learning and Molecular Mechanics Potentials

Towards chemical accuracy for alchemical free energy calculations with hybrid physics-based machine learning / molecular mechanics potentials

Tell machine learning potentials what they are needed for: Simulation-oriented training exemplified for glycine

Machine-Learning-Assisted Free Energy Simulation of Solution-Phase and Enzyme Reactions

Neural Network Potential with Multi-Resolution Approach Enables Accurate Prediction of Reaction Free Energies in Solution

Machine learning model for non-equilibrium structures and energies of simple molecules

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

Refining Potential Energy Surface through Dynamical Properties via Differentiable Molecular Simulation

Machine learning potentials with Iterative Boltzmann Inversion: training to experiment

Global Neural Network Potential with Explicit Many-Body Functions for Improved Descriptions of Complex Potential Energy Surface.

A Hessian-Based Assessment of Atomic Forces for Training Machine Learning Interatomic Potentials

Active learning meets metadynamics: Automated workflow for reactive machine learning potentials

Machine Learning of Reactive Potentials

Tutorial: How to Train a Neural Network Potential

Using machine learning to go beyond potential energy surface benchmarking for chemical reactivity

Beyond potential energy surface benchmarking: a complete application of machine learning to chemical reactivity