Amorphous materials — glasses, disordered solids — lack the periodicity that makes crystal structure determination straightforward. Their atomic arrangements are described statistically, not exactly. Determining the structure from spectroscopic measurements requires sampling an enormous space of possible configurations.
Note (2026-11-15; checked against the paper, v1): the paper's abstract says a score-based model learns a structural prior from low-fidelity data and then samples structures conditioned on differentiable spectral targets, so the model is a prior over structures steered by the measured spectrum, not a learned map from spectrum to structure as the essay words it. The paper itself calls structure determination an inverse problem, notes that meaningfully distinct structures can explain the same spectrum, and says its pair distribution function reconstructions were the most informative of the probes tested; it applies the method to amorphous silicon, a liquid-liquid transition in sulfur and ball-milled amorphous ice. The essay's contrast that the problem is generative and not inverse is the essay's framing.
The authors (arXiv:2603.23210) build GLASS, a generative framework that reconstructs realistic atomic structures directly from spectroscopic data. Given a measured spectrum, GLASS generates candidate structures whose simulated spectra match the measurement. No knowledge of the potential energy surface is needed — the model learns the spectrum-to-structure mapping from training data.
The through-claim: for amorphous materials, the structure determination problem is generative, not inverse. There is no unique structure corresponding to a spectrum — there's a distribution of structures consistent with it. GLASS samples from this distribution rather than attempting to find a single answer. The question shifts from “what is the structure?” to “what structures are consistent with this measurement?” — and the answer is a generative model, not a point estimate.