LDA-based machine learning force field for accurate electric-field-driven ferroelectric response in BaTiO3

Accurately modeling the electric-field response of BaTiO3 (BTO) requires a faithful reproduction of its equilibrium lattice structure, because dielectric permittivity and polarization switching are highly sensitive to subtle variations in tetragonality and Ti–O displacements. However, commonly used GGA functionals, such as PBEsol, systematically overestimate lattice constants, leading to significant underestimation of dielectric constants. In contrast, the local density approximation (LDA) provides structural parameters—particularly the c/a ratio and atomic off-centering—in much closer agreement with experimental data, offering a more reliable foundation for quantitative dielectric modeling. In this work, we construct a high-quality LDA-based first-principles dataset for BTO and fine-tune a machine learning force field (MLFF) together with a Born effective charge (BEC) prediction model. This integrated framework enables large-scale electric-field-coupled molecular dynamics simulations that retain the structural fidelity of LDA while achieving near–first-principles accuracy at significantly reduced computational cost. The resulting simulations successfully reproduce key ferroelectric features, including polarization switching and P–E hysteresis loops, and yield dielectric constants that are substantially closer to experimental values than those obtained from GGA-based approaches. These results demonstrate that adopting LDA as the foundational electronic structure method is essential for quantitatively accurate modeling of dielectric and ferroelectric responses in BTO, and that combining LDA-consistent MLFF and BEC models provides an efficient route to large-scale electric-field simulations of perovskite oxides.