DiffGA: Steering Frozen Diffusion Models for Materials Design via Evolutionary Search

Steering candidate generation pipelines toward user‑defined objectives is central to inverse materials design, yet it remains difficult when objectives are black‑box, ever-changing, or non‑differentiable. This is precisely the regime of traditional simulation‑ and experiment‑derived properties. Generative models offer a promising solution but have to be either retrained or fine‑tuned for every new objective, demanding large labeled datasets and substantial compute, or rely on differentiable score guidance that is brittle to gradient noise and undefined for many practical targets. We present DiffGA, a retraining‑free framework that injects evolutionary search directly into the sampling trajectory of a pretrained diffusion model, evolving partially denoised horizon states under an arbitrary black‑box oracle. Often campaign cost is dominated by oracle evaluation and, thus, we couple DiffGA with an on‑the‑fly Bayesian surrogate that reuses learned representations from state-of-the-art machine-learned force fields (MLFFs). A Gaussian process is fit during the campaign on the force field’s own graph embeddings and — once it demonstrably ranks an unseen generation — replaces the expensive property‑evaluation step under an uncertainty‑aware acquisition rule, cutting oracle cost with negligible loss of accuracy. Across three crystal‑design tasks using a frozen MatterGen backbone, DiffGA outperforms diffusion‑only, compute‑matched, purely evolutionary, and fine‑tuned baselines in the (oracle‑assessed) rate of stable, unique, and novel hits. Rather than proposing specific new compounds, we use density functional theory to quantify the fidelity of the MLFF oracle itself, showing that it is a faithful proxy in key regimes and systematically biased in others, delineating precisely when proxy‑driven generative discovery can be trusted. On a representative sample, DiffGA’s top hits survive first‑principles re‑evaluation as genuine, physically realistic hits, and the surrogate‑accelerated pipeline preserves this fidelity end‑to‑end at a fraction of the cost.