Robot laser-heats thin films while software picks each next heating condition

The work

Technical · from the record

SARA-H couples robotic lateral-gradient laser spike annealing of thin films with on-the-fly X-ray diffraction analysis: non-negative matrix factorization and the CrystalShift probabilistic labeling algorithm turn each annealed stripe into expected phase fractions, which update a Gaussian process model of the processing space, and a logEI acquisition function selects the next anneal condition. A human expert can add or remove candidate phases during a campaign, which changes the objective the active learning cycle optimises. The framework was run on Bi2O3, SnOx and a Bi-Ti-O composition spread; the Bi-Ti-O campaign is reported over 120 iterations divided into 45 exploration, 25 refined exploration and 50 targeted exploitation anneals. The resulting processing phase diagrams map extensive domains of delta-Bi2O3 and Bi2Ti2O7 and regions of Bi-substituted anatase, which the authors read as evidence that Bi doping inhibits the anatase-to-rutile transformation, and no sillenite Bi12TiO20 was observed.

How AI was used

Each iteration of the closed loop begins with an lg-LSA anneal of an amorphous thin film at a condition expressed as laser power and scan velocity, converted to peak temperature and dwell time through a thermoreflectance calibration. Spatially resolved synchrotron diffraction patterns from across the stripe are integrated with PyFAI, then reduced by a non-negative matrix factorization based on an extreme-ray finding algorithm (rank k=4), whose bases are fitted with a kernel ridge regression background model so that amorphous bases can be discarded. The retained bases are passed to CrystalShift, which performs a best-first tree search with lattice refinement and Laplace-approximated Bayesian model comparison over a user-supplied candidate phase set, with a temperature-scaling parameter calibrated on 10,000 synthetic XRD patterns; the top candidates per basis are combined with the NMF activation coefficients into expected phase activations as a function of temperature. These activations, with measurement and temperature uncertainties propagated, train a Gaussian process regression model with a Matérn-5/2 kernel over composition, log dwell time and peak temperature, which serves as the surrogate for Bayesian optimisation. A logEI acquisition function, optimised with BFGS and optionally summed over the temperature range of a stripe, selects the next anneal; the targeted phase is either held fixed (target sampling) or cycled round-robin over the candidate pool (cycle sampling). Human experts inspect the evolving phase fields mid-campaign and add or remove candidate phases, shifting the objective between exploration and exploitation. Sampling strategies were also compared on a synthetic model built from GPs fitted to hand-labeled Bi-Ti-O data, with random and uncertainty sampling as baselines.