Doubly Polarized QM/MM with Machine Learning Chaperone Polarizability.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Molecular Mechanics

Background:

• Semiempirical (SE) molecular orbital methods significantly underestimate molecular polarizability.
• This underestimation leads to errors in QM/MM simulations of solution-phase reactions, particularly in free energy profiles.
• Inadequate electronic polarization response to solvent fields is a key issue.

Purpose of the Study:

• To develop a hybrid framework that enhances the response property of SE/MM methods.
• To address the underpolarization problem in free energy simulations.
• To improve the accuracy of QM/MM methods for condensed-phase reactions.

Main Methods:

• Introduced a hybrid framework with "chaperone polarizabilities" on QM atoms.
• Machine learning (ML) determined chaperone polarizabilities to match condensed-phase AI polarizability.
• Developed the doubly polarized QM/MM (dp-QM/MM) method, combining SE wave functions and classical polarizabilities.
• Applied the method to free energy simulations of the Menshutkin reaction in water using AM1/MM.

Main Results:

• ML chaperones reduced solute molecular polarizability error from 6.78 to 0.03 ų compared to DFT.
• Chaperone correction added ~10 kcal/mol of polarization energy in the product region.
• Simulated free energy profiles showed closer agreement with experimental results.
• Modified free energy profiles via enhanced solvation corrections, particularly for charge-separated states.

Conclusions:

• The dp-QM/MM method effectively remedies the underpolarization issue in SE/MM free energy simulations.
• ML-driven chaperone polarizabilities provide a physically accurate correction.
• This approach enhances the reliability of QM/MM simulations for complex chemical processes.