[Paper Review] Neural Network Based in Silico Simulation of Combustion Reactions

Jinzhe Zeng, Liqun Cao+3 more|arXiv (Cornell University)|Nov 27, 2019

Machine Learning in Materials ScienceMaterials Science85 references21 citations

TL;DR

This paper presents a deep neural network potential trained on DFT data to enable high-accuracy, nanosecond-scale reactive molecular dynamics simulations of combustion reactions in large systems (hundreds of atoms). The method achieves DFT-level precision in energy and forces while being orders of magnitude faster than conventional DFT, enabling detailed mechanistic insights into H2 and CH4 combustion without predefined reaction coordinates.

ABSTRACT

Understanding and prediction of the chemical reactions are fundamental demanding in the study of many complex chemical systems. Reactive molecular dynamics (MD) simulation has been widely used for this purpose as it can offer atomic details and can help us better interpret chemical reaction mechanisms. In this study, two reference datasets were constructed and corresponding neural network (NN) potentials were trained based on them. For given large-scale reaction systems, the NN potentials can predict the potential energy and atomic forces of DFT precision, while it is orders of magnitude faster than the conventional DFT calculation. With these two models, reactive MD simulations were performed to explore the combustion mechanisms of hydrogen and methane. Benefit from the high efficiency of the NN model, nanosecond MD trajectories for large-scale systems containing hundreds of atoms were produced and detailed combustion mechanism was obtained. Through further development, the algorithms in this study can be used to explore and discovery reaction mechanisms of many complex reaction systems, such as combustion, synthesis, and heterogeneous catalysis without any predefined reaction coordinates and elementary reaction steps.

Motivation & Objective

• To develop a machine learning potential that achieves DFT-level accuracy for combustion reactions.
• To enable long-timescale reactive molecular dynamics simulations (nanoseconds) for large systems (hundreds of atoms) using neural network potentials.
• To explore complex reaction mechanisms in combustion, synthesis, and catalysis without relying on predefined reaction coordinates or elementary steps.
• To construct reference datasets from DFT calculations to train robust, transferable neural network potentials.

Proposed method

• Two reference datasets were constructed from high-level DFT calculations for H2 and CH4 combustion systems.
• Graph neural networks were trained to predict potential energy and atomic forces with DFT accuracy.
• The neural network potentials were validated against DFT data for energy and force predictions.
• Reactive molecular dynamics simulations were performed using the trained neural network potentials at scale.
• The simulations were run for nanosecond timescales to capture complex reaction pathways.
• The method operates without predefined reaction coordinates, allowing discovery of unknown mechanisms.

Experimental results

Research questions

• RQ1 Can a neural network potential trained on DFT data achieve DFT-level accuracy for combustion reactions?
• RQ2 Can such a potential enable nanosecond-scale reactive MD simulations of large systems (hundreds of atoms)?
• RQ3 Can the method uncover detailed reaction mechanisms in H2 and CH4 combustion without prior assumptions on reaction coordinates?
• RQ4 To what extent can this approach be generalized to other complex reaction systems like catalysis or synthesis?

Key findings

• The neural network potentials achieved DFT-level accuracy in predicting potential energy and atomic forces.
• The method enabled reactive molecular dynamics simulations of hundreds of atoms over nanosecond timescales, which is computationally infeasible with standard DFT.
• Detailed combustion mechanisms for H2 and CH4 were successfully revealed, including key intermediates and transition states.
• The approach does not require predefined reaction coordinates, allowing for the discovery of complex, non-trivial reaction pathways.
• The trained models demonstrated transferability and robustness across different system sizes and configurations.