An Accurate and Efficient Machine-Learned Potential for SiC from Ambient to Extreme Environments

TL;DR

The study addresses the need for an accurate yet computationally efficient potential to model SiC under extreme environments. The authors develop tabGAP, a DFT-trained Gaussian approximation potential, enabling multimillion-atom MD across ambient to extreme conditions and allowing comprehensive mapping of the pressure–temperature phase diagram, threshold displacement energy distributions, and large-scale collision cascades for SiC polymorphs. Key contributions include a diverse training dataset, validation against DFT and phonon data, a complete P–T phase diagram with incongruent melting behavior, and detailed radiation-damage characterization that reveals differences from traditional potentials. The work provides atomistic insight into phase transformations and radiation response in SiC and offers a transferable approach for extending ML-IAPs to other materials in extreme environments.

Abstract

Silicon carbide (SiC) polymorphs are widely employed as nuclear materials, mechanical components, and wide-bandgap semiconductors. The rapid advancement of SiC-based applications has been complemented by computational modeling studies, including both ab initio and classical atomistic approaches. In this work, we develop a computationally efficient and general-purpose machine-learned interatomic potential (ML-IAP) capable of multimillion-atom molecular dynamics simulations over microsecond timescales. Using the ML-IAP, we systematically map the comprehensive pressure-temperature phase diagram and the threshold displacement energy distributions for the 2H and 3C polymorphs. Furthermore, collision cascade simulations provide in-depth insights into polymorph-dependent primary radiation damage clustering, a phenomenon that conventional empirical potentials fail to accurately capture.

Figures (7)

• Figure 1: An Overview of the DFT calculated dataset and the validation and training accuracy of tabGAP. (a) The relationships among structures in the database are visualized via a two-dimensional embedding based on the SOAP similarity metric, with different polytypes highlighted in distinct colours: 3C (orange), 6H (red), 2H (blue), 4H (green), RS (brown), etc. A representative structure is shown for each polytype along with the fraction of the training data for each polytype (as a percentage of the total 185,542 atoms in the database). Note that the isolated Si and C atoms (not shown here) are also included in the database as a global reference for the potential. (b) Comparison of the equations of state from DFT and tabGAP for the five experimentally identified 2H/3C/4H/6H/RS polymorphs. Scatter plots of (c) energies and (d) force components versus DFT data.
• Figure 2: Phonon dispersion. Predicted phonon dispersions of the 3C, 2H, and RS SiC polymorphs compared to DFT calculations and (for 3C) experiments serrano2002determination. For the high-pressure RS phase, the calculations are done at 0 GPa and 70 GPa.
• Figure 3: Amorphous structures with applied 5% compressive strain (upper panel, corresponding to $\sim$45 GPa), no strain (middle panel, 0 GPa) and 5% tensile strain (bottom panel, $\sim$35 GPa). (a) Radial distribution functions (RDFs) of the amorphous configuration under various strains. (b) Bond angle distributions for Si–C bonds with a cutoff distance of 1.7 Å, which corresponds to the first minimum of the RDF.
• Figure 4: (a) Pressure-temperature phase diagram of SiC. The background shading in different colors emphasizes the stability regions of different phases. ML-tabGAP MD simulation results (denoted by black crosses), experimental data (red circles) miozzi2018equationtracy2019n, and DFT results (black squares) ran2021phaseivashchenko2019temperaturelee2015first are shown. The inferred phase boundaries are drawn with black dashed lines, and the shaded region denotes the metastable decomposition region. (b) Representative atomic snapshots at key phase boundary locations.
• Figure 5: Total energy difference for quasi-static simulations of (a) 2H- and (b) 3C-SiC using tabGAP, Tersoff/ZBL and DFT methods. Left panel: Schematic representations of four representative atomic displacement directions. Middle panel: Total energy differences during silicon atom displacement. Right panel: Total energy differences during carbon atom displacement.