Interatomic potentials which accurately describe long-range electrostatics require atom-centered charges. One such method to determine these atom-centered charges from density functional theory (DFT) calculations is the density-derived atomic point (DDAP) charge method. DDAP fits atom-centered Gaussians to the ground-state DFT charge density and preserves the multipole moments that govern long-range electrostatics. While these charges accurately predict long-range behavior, in practice, they are limited by their reliance on fixed, heuristic parameters and a constrained solver that becomes numerically unstable for complex or covalent systems. In this work, we present opt-DDAP, which solves this limitation by reformulating the algorithm as a differentiable computational graph. This reformulation allows for the optimization of Gaussian basis parameters and the reciprocal-space cutoff using automatic differentiation. To ensure numerical robustness through this automatic differentiation process, we replace the conventional Lagrange-multiplier approach with a pseudoinverse solution followed by charge renormalization, maintaining stability even in the presence of ill-conditioned matrices. We validate the framework on NaCl vacancy supercells and on MoS2, demonstrating faithful reconstruction of both absolute and difference charge densities. The optimized charges are intended to serve as inputs to effective electrostatic models in machine learning and empirical interatomic potentials that incorporate long-range interactions.
M. H, Sudarshan Vijay· Journal of Chemical Theory a...· 0 citations
Implementing novel features and experimental algorithms into widely adopted density functional theory (DFT) codes is frequently hindered by complex legacy architectures and the use of compiled languages such as Fortran. These production codes, while optimised for high-performance computing clusters, present significant hurdles for software development and rapid prototyping, often requiring deep expertise in the code's internal structure to modify. To address this challenge, we present a Python plugin infrastructure for the Vienna ab-initio Simulation Package (VASP) that combines computational efficiency with the flexibility of high-level scripting. Our architecture uses a C++ intermediate layer and pybind11 to expose VASP data as NumPy arrays via shared memory buffers, ensuring high performance without data duplication. We implement two categories of plugins: those that modify quantities at the end of each converged self-consistent field (SCF) cycle, such as structure and force_and_stress, and those that operate during the SCF cycle, such as local_potential and occupancies. We demonstrate the utility of our implementation through three applications, structure relaxation using the scipy library, implementing an implicit solvent model, and adding the DFT-D4 dispersion corrections. This infrastructure effectively bridges the gap between high-performance electronic structure routines and the widespread scientific Python ecosystem.
Sudarshan Vijay, M. Marsman, Georg Kresse et al.· 0 citations