Training Neural Nets To Learn Reactive Potential Energy Surfaces Using Interactive Quantum Chemistry in Virtual...

Area of Science:

• Computational Chemistry
• Machine Learning
• Data Science

Background:

• Computational workflows are increasingly limited by data curation challenges (size, quality, bias, format, coverage) rather than processing power.
• Machine learning (ML) methods, particularly neural networks (NNs), require high-quality data for accurate modeling of chemical processes.
• Traditional data generation methods for reactive potential energy surfaces (PESs) can be time-consuming and may not adequately sample important regions.

Purpose of the Study:

• To introduce an open-source, GPU-accelerated neural network (NN) framework for learning reactive potential energy surfaces (PESs).
• To investigate interactive ab initio molecular dynamics in virtual reality (iMD-VR) as a novel data curation strategy for training NNs.
• To compare the performance of NNs trained with iMD-VR data against those trained with traditional constrained molecular dynamics (MD) data.

Main Methods:

• Developed an open-source, GPU-accelerated neural network (NN) framework.
• Employed real-time interactive ab initio molecular dynamics in virtual reality (iMD-VR) for rapid sampling of reaction pathways.
• Generated training data for NNs using both iMD-VR and traditional constrained MD methods, focusing on CN radical hydrogen abstraction reactions with isopentane.

Main Results:

• Both iMD-VR and constrained MD trained NNs qualitatively reproduced key features of the reactive PESs, including low and early reaction barriers.
• NN performance is sensitive to the training data set; iMD-VR excels at sampling near the minimum energy path (MEP).
• NNs trained on iMD-VR data accurately predict energies near the MEP but are less accurate for 'off-path' structures, while constrained MD NNs perform better on high-energy off-path structures.

Conclusions:

• Interactive virtual reality molecular dynamics (iMD-VR) is an effective strategy for curating training data for neural networks (NNs) in computational chemistry.
• The choice of data curation strategy significantly impacts the accuracy and predictive capabilities of NNs for learning potential energy surfaces (PESs).
• Future work should consider hybrid data curation approaches to balance sampling near the minimum energy path and high-energy off-path regions for improved NN performance.