Does Hessian Data Improve the Performance of Machine Learning Potentials?

Area of Science:

• Computational chemistry
• Materials science
• Drug discovery

Background:

• Machine Learning Interatomic Potentials (MLIPs) predict energies and forces with quantum accuracy.
• Force fitting in MLIP training improves potential-energy surface predictions.
• Hessian matrix training encodes second-order information about the potential energy surface (PES) curvature.

Purpose of the Study:

• To evaluate the integration of Hessian matrix training in MLIPs.
• To assess the impact of Hessian training on extrapolation to nonequilibrium geometries.
• To analyze the benefits and limitations of Hessian integration in MLIPs for various computational chemistry applications.

Main Methods:

• Training MLIPs with varying combinations of energy, force, and Hessian data.
• Evaluating model performance on equilibrium and first-order saddle point geometries.
• Testing extrapolation capabilities to nonequilibrium geometries using a small molecule reactive dataset.

Main Results:

• Hessian-trained MLIPs demonstrate improved extrapolation to unseen molecular systems.
• Hessian integration enhances accuracy in reaction pathway modeling and vibrational spectra prediction.
• Hessian training can reduce the total data required for effective MLIP models, despite increased computational expense.

Conclusions:

• Hessian integration in MLIPs offers significant advantages for predicting molecular properties and reaction dynamics.
• Practitioners should weigh the benefits of improved accuracy and data efficiency against the increased computational cost.
• This work provides insights for informed decisions on employing Hessian-trained MLIPs in computational chemistry research.