Scaling Density Functional Theory with Gaussian Splatting
GS-DFT claims conventional large-basis accuracy with fewer parameters, scaling to 2,742 atoms on one four-GPU node.
The paper represents molecular orbitals as optimized clouds of Gaussians, borrowing the splatting idea but replacing rendering with quantum-mechanical energy minimization. Its solver adds adaptive density fitting with screening and differentiable orbital orthogonalization. In tests, the optimized basis converged in energy, density, and nuclear forces, and handled stretched bonds and anions at equal parameter counts where fixed bases needed augmentation. ArXiv · AI/CL/LG's note
The paper represents molecular orbitals as optimized clouds of Gaussians, borrowing the splatting idea but replacing rendering with quantum-mechanical energy minimization. Its solver adds adaptive density fitting with screening and differentiable orbital orthogonalization. In tests, the optimized basis converged in energy, density, and nuclear forces, and handled stretched bonds and anions at equal parameter counts where fixed bases needed augmentation. ArXiv · AI/CL/LG's note
score 5