GO-Diff: Data-free and amortized global structure optimization

We introduce GO-Diff, a diffusion-based method for global structureoptimization that learns to directly sample low-energy atomicconfigurations without requiring prior data or explicitrelaxation. GO-Diff is trained from scratch using a Boltzmann-weightedscore-matching loss, leveraging only the known energy function toguide generation toward thermodynamically favorable regions. Themethod operates in a two-stage loop of self-sampling and modelrefinement, progressively improving its ability to target low-energystructures. Compared to traditional optimization pipelines, GO-Diffachieves competitive results with significantly fewer energyevaluations. Moreover, by reusing pretrained models across relatedsystems, GO-Diff supports amortized optimization — enabling fasterconvergence on new tasks without retraining from scratch.