A learning algorithm ran a synchrotron beamline to find a phase-change material

The work

Technical · from the record

CAMEO is a closed-loop active-learning system that took real-time control of high-throughput X-ray diffraction at a synchrotron beamline to map structural phases of a Ge–Sb–Te composition spread and search it for a phase-change material with the largest difference in optical bandgap between amorphous and crystalline states. A graph-based Bayesian phase-mapping model, label propagation to unmeasured compositions, and per-phase-region Gaussian process regression chose each next composition to measure, while a human expert fitted the raw ellipsometry data to return the bandgap difference for that composition. The run identified a composition at a phase boundary, averaging Ge4Sb6Te7, over 19 iterations taking approximately 10 h, compared with over 90 h to measure all 177 composition spots; its ΔEg was 0.76 ± 0.03 eV against 0.23 ± 0.03 eV for Ge2Sb2Te5. Electron microscopy of this composition showed a coherent nanocomposite of GST and Sb–Te phases, and photonic switching devices made from it were stable over at least 30,000 cycles.

How AI was used

Raw ellipsometric spectra collected before the run were reduced to pairwise similarity weights that modified the edges of a composition graph, supplying a prior for phase mapping. At each cycle the measured diffraction patterns were segmented into phase regions using GRENDEL, a Markov-random-field, graph-based phase-region and constituent-phase identification method that encodes physical constraints such as the Gibbs phase rule. Gaussian random field harmonic energy minimization, a semi-supervised method, propagated phase-region labels and their likelihoods to compositions not yet measured, and Gaussian process regression propagated the functional property. Phase-mapping measurements were chosen by risk minimisation over the resulting hypothesis space of phase maps until the Fowlkes–Mallows index between iterations reached 80% convergence, after which a separate Gaussian process was fitted to each phase region, the top-ranked region selected, and the next composition chosen by an upper-confidence-bound acquisition function with an added term weighting distance from the phase boundary. Each selected composition was then measured by synchrotron diffraction under programmatic instrument control and passed to a human expert who fitted Drude and Tauc–Lorentz models to the raw ellipsometry data and returned the bandgap difference to the loop. The scheme was compared post hoc against GP-UCB and random sampling over 100 simulated runs, and hyperparameters had been tuned beforehand on a previously characterised Fe–Ga–Pd spread.