A deep neural network potential for carbon
Abstract
Carbon materials and their unique properties have been extensively studied by molecular dynamics. While the potentials used to model the interactions between carbon atoms have become increasingly more accurate over the years, they still have their limitations. These mostly stem from the fixed mathematical expressions used to construct the potentials and limits their application to structures similar to those the potential was constructed for. Machine learning (ML) has recently emerged as a way of addressing some of the limitations of these classical potentials, where ML is used to train a potential using highly accurate reference data obtained from first-principle calculations. The trained potentials can then infer the energy (and by the derivative also the forces) of a system of atoms and can thus be used to perform molecular dynamics.
For modeling of carbon materials, the recently published Gaussian approximation potential (GAP-20) by Rowe et al. [1] is considered as the current state-of-the-art. However, GAP-20 is not without its faults. A recent study investigating the accuracy of GAP-20 found that the predicted van der Waals interactions are inaccurate [2]. Furthermore, Gaussian approximation potentials are based on kernel ridge regression and are thus by their nature limited in the amount of training data that can be used and their inference speed is slow compared to neural networks. This project is aimed at solving the two major limitations of GAP-20 (the inaccuracy in the van der Waals interactions and the slow inference speed) by training deep neural network potentials to predict energies and forces. Using the same GAP-20 dataset the training will be performed using the DeePMD toolkit [3]. This includes developing the neural network architecture, optimization of hyperparameters and if necessary, also expanding the GAP-20 dataset with additional structures to improve the accuracy of the van der Waals interactions. Any additional training data will be generated using highly accurate first-principles density-functional theory calculations performed using VASP. DeePMD also interfaces with the molecular dynamics package LAMMPS which will be used to verify the accuracy of the predicted energies. We expect that our results will exceed the current state of the art ML potential used to model carbon materials (GAP-20). With improvements of both the speed and accuracy it can become the de facto potential for researchers to use when modeling carbon materials at large scales.
References
[1] P. Rowe, V. L. Deringer, P. Gasparotto, G. Csányi and A. Michaelides, “An accurate and transferable machine learning potential for carbon”, J. Chem. Phys., vol. 153, nr. 3, pp. 034702, July 15, 2020. DOI: 10.1063/5.0005084.
[2] C. Qian, B. McLean, D. Hedman and F. Ding, “A comprehensive assessment of empirical potentials for carbon materials”, APL Materials, vol. 9, nr. 6, pp. 061102, June 1, 2021. DOI: 10.1063/5.0052870.
[3] G. Csanyi, “Carbon GAP 20”, -07-01T10:42:01Z 2020. DOI: 10.17863/CAM.54529
