AMASE
Autonomous Materials Search Engine that closes the experiment–theory loop in real time by interleaving robotic thin-film X-ray diffraction with CALPHAD-based Gibbs-free-energy phase-diagram prediction to autonomously map binary phase diagrams.
| Affiliation | University of Maryland MSE and Maryland Quantum Materials Center (Takeuchi group), with NIST (Kusne, McDannald) |
| First introduced | 2024-10 (arXiv:2410.17430) |
| Lifecycle stages | Multi-stage (active-learning experiment selection, robotic XRD measurement, peak detection and classification, CALPHAD thermodynamic update) |
| Autonomy level | Fully autonomous — a single live run completes without human intervention beyond mounting the composition-spread wafer |
| Domain focus | Materials science — thin-film eutectic binary phase diagrams; demonstrated on Sn–Bi |
| Availability | Open source — data and code released as part of the manuscript |
Approach
AMASE runs threaded cyclical tasks that alternate measurement and thermodynamic computation:
- Active-learning composition selection. A variational Gaussian Process Classifier (VGPC) uses the current CALPHAD-evaluated phase boundary as its prior and selects the next composition to probe at a fixed temperature; convergence is signalled when the tracked Bi (012) or Sn (101) peak drops below noise on one side of a candidate boundary and is confirmed at ±0.01 in composition.
- XRD measurement and peak detection. A modified 1D YOLO object-detection model (adapted from computer vision) extracts peak positions, FWHMs, and intensities from each diffraction pattern; peaks are indexed against the Inorganic Crystal Structure Database.
- CALPHAD update. Once a phase-boundary composition is identified at a temperature, Thermo-Calc minimises Gibbs free energies of the three phases (β-Sn, Bi, liquid) using Redlich–Kister polynomials, run ten times to estimate uncertainty; the updated phase diagram is fed back as the prior for the next iteration.
- Next-temperature selection. A GP acquisition function in exploration mode chooses the next temperature within a 30 °C window above the current one, using uncertainty in the calculated solvus composition.
- End-point detection. A separate GP classifier detects vanishing of the solvus phase boundary and triggers a jump to the liquidus search routine.
The phase-boundary-search-routine averages six XRD iterations and ~40 min per temperature.
Validation
End-to-end live demonstration on the Sn–Bi thin-film binary system. AMASE constructed the thin-film phase diagram in a single 8 h 22 min autonomous run consisting of 66 XRD measurements across 11 temperature points, deliberately seeded only with a 0.71 < x < 0.95 Sn-rich composition spread and the bulk phase-diagram free-energy functional forms. The predicted eutectic point was independently validated with a separate thin-film spread sample focused near x = 0.5.
Notable results
- AMASE-predicted eutectic point at Sn 53.3 % ± 2 % and 133.1 °C ± 1 °C agreed within 3 % with the independently measured 55.5 % ± 1.5 %, 133.2 °C ± 10 °C — and deviated from the bulk eutectic (59.5 %, 140.7 °C) as expected for a thin film.
- 6-fold reduction in experiments versus an exhaustive 10 °C × composition grid (66 vs ~400 XRD measurements; 8 h vs >60 h continuous run), critical for volatile Sn–Bi films that would oxidise during longer thermal exposure.
- Ablation against a GP-only navigation workflow (with post-hoc CALPHAD extrapolation): over 20 simulated runs, AMASE used significantly fewer measured points with tighter distributions, attributed to CALPHAD’s ability to extrapolate phase-diagram features that GP alone cannot skip over.
Primary paper
Other references
None yet.
Code
Data and code released as part of the manuscript (“Data that supports the finding of this study as well as the code have been made available as a part of the manuscript”).