AI Model Reconstructs Molecular Structures From X-Ray Blast Data
Researchers at SLAC National Accelerator Laboratory built a machine learning model that predicts the 3D structure of molecules by analyzing how their ions scatter after being hit with powerful X-rays. The work, published in Nature Communications, demonstrates a practical method for imaging individual molecules during chemical reactions-a capability that could advance drug discovery and industrial chemistry.
The technique, called Coulomb explosion imaging, works by firing an X-ray pulse at a single molecule in a vacuum. The pulse strips away electrons, leaving positively charged ions that repel each other violently and strike a detector. The detector records the momentum of each ion, creating a dataset that can theoretically be reversed to reveal the molecule's original structure.
The Computing Problem
Reconstructing molecular geometry from ion momentum has remained impractical for decades. The ions don't explode instantaneously-there's a brief delay during which atoms shift slightly, making simple physics equations unreliable. Each additional atom in the molecule multiplies the computational difficulty exponentially.
"It's kind of like breaking a glass and trying to put it back together from how the pieces flew apart," said Phay Ho, a physicist at Argonne National Laboratory and co-author of the study.
Training on Two Datasets
The research team, led by Xiang Li at SLAC's Linac Coherent Light Source, developed a generative AI model called MOLEXA (molecular structure reconstruction from Coulomb explosion imaging). Rather than solving equations directly, the model learns patterns from training data to make statistical predictions.
The team created training data using a quantum mechanics simulation that ran for over a month, producing 76,000 molecular samples. When the model trained on this dataset alone, predictions were inaccurate. The researchers then added a second, larger dataset generated using classical physics equations-less precise but roughly 100 times bigger.
This two-step training approach cut prediction errors in half, according to Li.
Testing Against Known Structures
The researchers tested MOLEXA using experimental data from the European X-ray Free-Electron Laser facility in Germany. They reconstructed molecules including water, tetrafluoromethane, and ethanol, then compared the results to established structures from the National Institute of Standards and Technology.
The model's predictions largely matched known structures. Chemical bonds appeared in the correct locations, with only minor variations in angles. Position errors generally stayed below half the length of a typical chemical bond.
Next Steps: Larger Molecules and Real-Time Reactions
The current model handles molecules with fewer than ten atoms. The team plans to scale it to larger systems and apply it to time-resolved experiments that capture molecules mid-reaction-essentially creating molecular movies that show how chemical transformations unfold.
The researchers are also testing whether the model can reconstruct molecules when the detector misses some ions. If successful, the technique could eventually apply to proteins, which contain thousands of atoms and have direct relevance to biology and medicine.
James Cryan, interim deputy director for science at LCLS, said the work will help interpret data from SLAC's superconducting X-ray laser, which delivers X-ray pulses at high rates.
The collaboration included researchers from Stanford PULSE Institute, Stanford University, Kansas State University, and institutions in Germany and France. The Department of Energy's Office of Science provided primary funding.
Learn more: AI for Science & Research | AI Learning Path for Research Scientists
Your membership also unlocks:
