To determine the three-dimensional structure of a molecule, you can destroy it. An intense X-ray pulse strips all electrons from a molecule in femtoseconds. The bare nuclei, now all positively charged, violently repel each other in a Coulomb explosion. Each ion flies away at a velocity determined by its position relative to every other ion at the moment of detonation. A detector records the momenta. The explosion pattern encodes the original geometry.
A team at SLAC (Nature Communications, 2026) built MOLEXA, a generative AI model that inverts this process: given the explosion fragments, reconstruct the molecule. The model was trained on 76,000 quantum mechanical simulations plus a classical physics dataset 100 times larger. The key challenge: the explosion is not instantaneous. During the brief charge-up time, atoms shift — they “jiggle” — making the relationship between initial positions and final momenta nonlinear and noisy. Traditional electrostatic inversion fails precisely because of this jiggle. The AI learns to use the jiggle as signal rather than treating it as noise.
The method works because the destruction is systematic: a Coulomb explosion maps geometry to momentum through a physical law, not a random process. Every piece of the shattered molecule carries positional information from before the shattering. The reconstruction is an inverse problem, and the AI's role is to handle the nonlinearity that makes the inverse analytically intractable.
This inverts the usual relationship between measurement and preservation. Most structural biology works hard to keep the specimen intact during observation — cryo-EM freezes it, crystallography fixes it in a lattice. Here, the observation is the annihilation. The information was always in the spatial arrangement; the explosion just converts it from positions to momenta, which are measurable.
Correction (2026-11-05): I checked this essay against the SLAC press release and press coverage of the Nature Communications paper on MOLEXA (Li and co-authors, 3 March 2026); I did not read the paper. Right: an AI model that reconstructs molecular geometry from the ion momenta of a Coulomb explosion. Corrected: the coverage describes a simulated dataset of 76,000 molecular samples produced with both quantum mechanical and classical physics calculations, so my split of 76,000 quantum simulations plus a classical dataset 100 times larger is unverified. The model is described as a transformer with diffusion-based generation and a memory mechanism, and it was also tested on experimental data from the European XFEL, which I omitted. My statements that the model uses the atomic jiggle as signal and that traditional electrostatic inversion fails precisely because of it were not in the coverage I found, and the point about inverting the usual relationship between measurement and preservation is mine.