Transferability of MACE Graph Neural Network for Range Corrected Δ-Machine Learning Potential QM/MM Applications

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Quantum mechanics/molecular mechanics (QM/MM) simulations

Background:

• The accuracy of QM/MM simulations is crucial for understanding complex chemical reactions.
• Previous range-corrected Δ-machine learning potentials (ΔMLP) improved QM/MM accuracy by correcting energies and forces.
• Deep neural networks have shown promise in enhancing simulation accuracy.

Purpose of the Study:

• To extend the ΔMLP approach by incorporating graph neural networks, specifically the MACE architecture.
• To evaluate the transferability and accuracy of AM1/d + MACE models for phosphoryl transesterification reactions.
• To compare the performance of MACE against the DeepPot-SE (DP) architecture in QM/MM simulations.

Main Methods:

• Training AM1/d + MACE models to reproduce PBE0/6-31G* QM/MM energies and forces.
• Testing model transferability using reactions not included in the training set.
• Calculating free energy surfaces to assess reaction pathway accuracy.
• Varying MACE hyperparameters to study their impact on accuracy and performance.

Main Results:

• AM1/d + MACE models accurately reproduced target free energy surfaces, outperforming AM1/d + DP models in some cases.
• End-state AM1/d + MACE models correctly predicted a stable pentacoordinated phosphorus intermediate, even without similar structures in training data.
• MACE architecture demonstrated improved transferability for ΔMLP models.
• AM1/d + MACE simulations were found to be 28% slower than AM1/d QM/MM when using GPU acceleration.

Conclusions:

• The MACE architecture offers improved transferability for ΔMLP models in QM/MM simulations.
• Graph neural networks represent a promising direction for developing more accurate and transferable machine learning potentials.
• The developed AM1/d + MACE models provide a more reliable approach for studying complex chemical reactions like phosphoryl transesterification.