Self-Parametrizing System-Focused Atomistic Models.

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Accurate atomistic models are crucial for simulating chemical reactions in complex environments like proteins and nanostructures.
• Existing parametrization methods are often system-specific and lack automation, hindering broad applicability.

Purpose of the Study:

• To develop a fast, automated, and reliable parametrization procedure for atomistic models applicable to diverse nanoscale systems.
• To integrate quantum chemically derived molecular mechanics with machine learning for enhanced accuracy and uncertainty quantification.

Main Methods:

• A partial Hessian fitting procedure generates an initial physically motivated model from energy and Hessian data.
• A Δ-machine learning model is trained on on-the-fly evaluated reference data for energy and force corrections, including uncertainty estimates.
• A fragmentation approach enables the parametrization of large systems, with modular steps allowing for continuous model improvement.

Main Results:

• The developed approach provides an automatically parametrizable molecular mechanics model with machine-learned corrections.
• The method incorporates autonomous uncertainty quantification and refinement, tailoring model flexibility to available data.
• The approach facilitates the generation of system-focused electrostatic embedding environments for quantum-mechanical/molecular-mechanical hybrid models.

Conclusions:

• This automated, system-focused parametrization strategy significantly enhances the efficiency and reliability of computational studies for nanoscale systems.
• The integration of machine learning with quantum chemically derived models offers a flexible and reproducible framework for complex molecular simulations.
• The methodology is adaptable for various applications, including the creation of hybrid quantum-mechanical/molecular-mechanical models for diverse atomistic structures.