Development and application of a high-dimensional neural network potential for boron carbide

Abstract A high-dimensional neural network potential (HDNNP) is developed to accurately model the potential energy surface of boron carbide. The HDNNP is trained on an extensive dataset of structures generated using a Monte Carlo algorithm, covering a wide range of stoichiometries and structural configurations. Density functional theory (DFT) calculations provided energy and atomic force data, which were encoded using atom-centered symmetry functions to capture local atomic environments. The resulting HDNNP exhibits exceptional predictive performance, accurately estimating energies and forces across diverse structural configurations of boron carbide. The model effectively captures essential structural, energetic, and mechanical properties, including elastic behavior and the influence of stoichiometry on stability. The HDNNP development followed an iterative refinement strategy, in which the initial training set was systematically expanded to 2931 structures by incorporating finite-temperature AIMD snapshots at 300 K, 600 K, and 1000 K, as well as extrapolative configurations from MD simulations and defect-containing cells. This approach resulted in an HDNNP model that is numerically stable and physically consistent up to 500 K, with only minor drift observed at 600 K. The HDNNP achieves exceptional accuracy in predicting energies, forces, and mechanical properties across temperatures, reproducing Young’s modulus values of 416 GPa at 10 K and 420 ± 20 GPa at 300 K, in close agreement with DFT and experimental ranges. The predicted ultimate tensile strength of 43 GPa at a failure strain of 0.14 at 300 K is consistent with atomistic simulations (40.23 GPa, strain 0.12). Comparative MD simulations show that the HDNNP outperforms Tersoff and ReaxFF in both accuracy and computational efficiency, enabling timesteps up to an order of magnitude larger, achieving 43% faster performance than Tersoff and 79% faster than ReaxFF. These results highlight the HDNNP as a robust and efficient tool for simulating boron carbide across a range of temperatures and deformation conditions, offering quantum-level accuracy for large-scale atomistic modeling in extreme environments.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.