Benchmark study on deep neural network potentials for small organic molecules.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Materials Science

Background:

• Machine learning (ML) potentials (NNPs) approximate potential energy surfaces (PES), accelerating chemical exploration and property prediction.
• NNPs circumvent the need for explicit electronic Schrödinger equation solutions, offering a faster alternative to quantum mechanical methods.
• Standardized comparative evaluations are lacking for novel ML methods, hindering their widespread adoption.

Purpose of the Study:

• To compare the accuracy and transferability of four selected neural network potentials (NNPs): ANI, PhysNet, SchNet, and BAND-NN.
• To evaluate the performance of these NNPs in representing the potential energy surfaces (PES) of small organic molecules.
• To assess the suitability of these models for molecular dynamics simulations and geometry optimization.

Main Methods:

• Comparative evaluation of four NNPs (ANI, PhysNet, SchNet, BAND-NN) on two distinct test sets of small organic molecules.
• Assessment of accuracy using root mean square error (RMSE) on small organic molecules and GDB-11 database samples.
• Evaluation of PES smoothness via scans for bond stretch, angle bend, and dihedral rotations; geometry optimization for minimum energy structures; and isomer differentiation.

Main Results:

• ANI and SchNet achieved RMSEs of 0.55 and 0.60 kcal/mol, respectively, on small organic molecules (up to 8 heavy atoms).
• On GDB-11 molecules (10 heavy atoms), ANI yielded an RMSE of 1.17 kcal/mol, while SchNet had an RMSE of 1.89 kcal/mol.
• All tested models accurately differentiated isomers and demonstrated potential for molecular dynamics simulations and geometry optimization.

Conclusions:

• ANI and SchNet demonstrate high accuracy and transferability for small organic molecules, making them suitable for computational chemistry applications.
• The evaluated NNPs show promise for accelerating molecular dynamics simulations and predicting molecular properties with quantum mechanical accuracy.
• Further development and standardization of comparative evaluations are crucial for advancing ML applications in computational chemistry.