Geometry Optimization with Machine Trained Topological Atoms.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Traditional molecular geometry optimization relies on bonded potentials.
• Developing accurate and efficient energy functions is crucial for computational chemistry.

Purpose of the Study:

• To present and validate a novel energy function, FFLUX, for molecular geometry optimization.
• To demonstrate the capability of FFLUX in predicting atomic energies for unseen molecular geometries.

Main Methods:

• FFLUX utilizes topologically-partitioned atomic energies trained via kriging.
• Machine learning (kriging) predicts Intermolecular Quantum Analysis (IQA) atomic energies.
• Geometry optimization of a water molecule was performed using the FFLUX method.

Main Results:

• FFLUX accurately optimized 2000 distorted geometries to within 0.28 kJ/mol of ab initio energies.
• 50% of optimizations achieved sub-0.05 kJ/mol accuracy.
• Kriging models demonstrated robustness, optimizing geometries outside the training data range.

Conclusions:

• FFLUX provides a viable alternative to traditional bonded potentials for geometry optimization.
• The method shows high accuracy and robustness, even with limited or out-of-distribution training data.
• Analysis of energy components revealed independent behavior, offering chemical intuition distinct from standard force fields.