GS-DFT: Gaussian Splatting scales DFT simulations to 2,742 atoms on 4 H200s
sarahdrinkwater · x · 2026-09-28
Researchers from Mila, Université de Montréal and Princeton propose GS-DFT, representing molecular orbitals as a cloud of Gaussians whose positions, shapes and mixing coefficients are optimized jointly by gradient descent to minimize energy — no training data needed. Conceptually, it's 3D Gaussian splatting with the renderer replaced by quantum mechanics.
- Two key solver components: adaptive density fitting with screening for two-electron integrals, and regularized differentiable orbital orthogonalization.
- The optimized basis matches the largest conventional basis sets with a fraction of the parameters, converging systematically in energy, density and forces.
- At equal parameter count it captures stretched-bond and anion physics that fixed bases only recover with specialized augmentation.
- Quadratic peak memory scaling enables systems of up to 2,742 atoms on four H200 GPUs.
The authors say the next frontier is reliable forces at lower compute cost, as adaptive bases push differentiable DFT to larger systems.
Related event: GS-DFT Scales Quantum Chemistry to 2,742 Atoms via Gaussian Splatting(4 posts)→
More from Research
- Pinductor uses LLM priors to learn POMDP world models with 330x fewer episodes than DreamerV3 — burny_tech · 2026-09-29
- MIT lab's PEM-UDE recovers interpretable chaos equations from noisy data — burny_tech · 2026-09-29
- SIGReg regularizer in LeJEPA and LeWM reveals hidden contrastive pairwise repulsion in JEPAs — burny_tech · 2026-09-29
- New article argues AI models are likely more capable than they appear — zetalyrae · 2026-09-29
- NanoGPT embedding table gains validated by earlier ngrammer paper, both using AdaGrad — _arohan_ · 2026-09-29
- AI-Generated Proofs Deserve Public Posting, Argues Researcher, Readers Can Just Ask AI to Rewrite Them — aran_nayebi · 2026-09-29