Multiscale Quantum Mechanics/Molecular Mechanics Simulations with Neural Networks

Area of Science:

• Computational Chemistry
• Biochemistry
• Molecular Modeling

Background:

• Multiscale quantum mechanics/molecular mechanics (QM/MM) methods are vital for studying chemical and biological processes.
• High computational costs of ab initio QM/MM limit their application to complex biochemical systems.
• Semiempirical QM/MM offers higher efficiency but requires accurate corrections to achieve ab initio QM/MM level accuracy.

Purpose of the Study:

• To develop a neural network (NN) based method to improve the accuracy of semiempirical QM/MM simulations.
• To enable prediction of potential energies at the ab initio QM/MM level using semiempirical QM/MM simulations.
• To accelerate the calculation of free energy changes for chemical reactions.

Main Methods:

• Developed an extension of the Behler-Parrinello neural network representation for QM/MM calculations.
• Integrated the NN potential energy predictions with semiempirical QM/MM simulations.
• Applied the method to three reactions in water to calculate free energy changes.

Main Results:

• The NN-corrected semiempirical QM/MM method accurately predicted free energy changes for reactions in water.
• Results showed excellent agreement with reference data from ab initio QM/MM simulations.
• Achieved a speed-up of 1-2 orders of magnitude compared to direct ab initio QM/MM corrections.

Conclusions:

• The developed neural network method combined with semiempirical QM/MM provides an efficient and reliable strategy for chemical reaction simulations.
• This approach significantly reduces the computational cost associated with achieving ab initio QM/MM accuracy.
• The method holds promise for broader applications in computational chemistry and biochemistry.