An Accurate Machine-Learning Interatomic Potential for Disordered Hard Carbon Materials

Hard carbon (HC) is a critically important nanoporous material for energy storage applications, such as batteries and supercapacitors. A fundamental understanding of its properties requires accurate modeling of its complex, disordered atomic structure. While molecular dynamics (MD) simulations are essential for this task, achieving a balance between computational efficiency and quantum-mechanical accuracy remains a significant challenge. Existing machine-learning potentials (MLPs) often fall short because they typically ignore the effect of heteroatoms (e.g., H, O), which are ubiquitous in experimental HC samples and crucial for determining their electrochemical properties. Furthermore, the performance of different MLP architectures in capturing the potential energy surface of such disordered systems has not been systematically benchmarked.

To address these limitations, we developed a novel MLP framework that explicitly incorporates impurity atoms. We constructed a representative dataset for HC by sampling structures with densities ranging from 1.5 to 2.0 g cm⁻³ at temperatures between 1300 and 2200 K. Using this dataset, we trained and rigorously compared several state-of-the-art MLP architectures. The resulting potential demonstrates high accuracy, as validated by simulating HCs with varying H/O ratios. The computed X-ray diffraction (XRD) spectra and pair distribution functions (PDF) show excellent agreement with experimental data. Our work establishes a robust new paradigm for developing accurate and transferable MLPs, enabling high-fidelity MD simulations of realistic hard carbon materials for energy storage.