(370f) Neural Network and Reaxff Comparison for Au Properties

Atomistic potentials, such as reactive force fields¹ (ReaxFF), have been in use for decades to approximate the complexities of the potential energy surface for various chemical systems. Such potentials are critical tools for studying large atomistic systems or molecular dynamic simulations. This is because the ab-initio tools used to train these potentials are often too computationally intensive for such applications. However, such atomistic potentials are necessarily less accurate for certain applications, since they have historically been trained to a limited amount of ab-initiodata.

Recently, Behler and Parrinello have developed means for training a neural network atomistic potential²which can operate on unit-cells without a fixed number of atoms. These machine learning based potentials have already proven effective for system with few chemical systems, but much work is left to be done improving these methods.

In this work we have compared training methods for both ReaxFF and Behlerâ??Parrinello neural network (BPNN) atomistic potentials on various single-component Au systems. We chose Au for this study because of its diversity of known nanoscale structures. The fact that long-range electronic interactions are screened in Au makes it an appropriate system to model with atomistic potentials that are less suited for long-range interactions such as ReaxFF or a single BPNN.

Both potentials were trained using subsets of 9,972 density functional theory calculations that included 905 bulk, 1,022 surface, and 8,045 cluster configurations. 848 of the total calculations are fully relaxed systems, while the remaining were relaxation steps from the initial guess of each image. Each method was used to generate multiple potentials which were validated against a set of untrained calculations.

The best performing ReaxFF potential was trained from the 848 fully relaxed images and could reliably predict surface and bulk data; however, it was substantially less accurate for molecular clusters of 126 atoms or fewer. Training the ReaxFF potentials to more data often resulted in overfitting and lower accuracy overall.

In contrast, the BPNN performance seemed only to improve with additional data. Thus, the best performing BPNN was trained to all calculations not reserved for validation. This potential performed comparably or better than ReaxFF across all regimes. However, the BPNN potential in this implementation brings significantly higher computational cost. It was also significantly slower to implement using the open-source tools available during this work.

For additional details, see published work:
Boes, Jacob R., Groenenboom, Mitchell C., Keith, John A., Kitchin, John R. "Neural network and ReaxFF comparison for Au properties." Int. J. Quant. Chem., 2 March 2016.

1. A. C. T. van Duin, S. Dasgupta, F. Lorant, W. A. Goddard, J. Phys. Chem. A 2001, 105, 9396.
2. J. Behler, M. Parrinello, Phys. Rev. Lett. 2007, 98, 146401.