Development and Evaluation of Geometry Optimization Algorithms in Conjunction with ANI Potentials.

Area of Science:

• Computational Chemistry
• Molecular Mechanics
• Quantum Chemistry

Background:

• Molecular mechanical force fields (MMFFs) require accurate torsional angle parameters, often derived from computationally expensive *ab initio* calculations.
• Active learning potentials (ANI-1x) offer an efficient and accurate alternative for generating potential energy surfaces (PESs).

Purpose of the Study:

• To develop and assess a computational protocol for MMFF parameterization using ANI-1x, focusing on constrained geometry optimization.
• To evaluate the performance of ANI-1x with various optimization algorithms for reproducing DFT energies and geometries.

Main Methods:

• Developed a computational protocol for constrained geometry optimization using force projection for Cartesian coordinates.
• Assessed ANI-1x with Conjugate Gradient (CG) and quasi-Newton algorithms (BFGS, L-BFGS) for geometry optimization.
• Performed large-scale assessments on drug molecules and organic molecules for MMFF development.

Main Results:

• Achieved an average root-mean-square error (RMSE) of 1.3 kcal/mol between ANI-1x and DFT for drug molecule optimizations.
• Generated torsional PESs with an RMSE of 1.23 kcal/mol for conformers.
• Demonstrated that CG algorithms with ANI-1x accurately reproduce DFT-optimized geometries and torsional PESs.

Conclusions:

• The developed computational protocol and algorithms are effective for MMFF parameterization.
• CG-BS and CG-WS are suitable for generating PESs, while CG-WS or BFGS are ideal for full geometry optimization.
• This work provides a foundation for improving small-molecule MMFFs with advanced computational methods.