Researchers at NTT have developed a self-driving laboratory system that extracts human-executable growth rules from autonomous material synthesis experiments, according to a study published in Nature Communications. The team applied the framework to sputtering epitaxy to synthesize beta-gallium oxide (β-Ga2O3) films.
The automated setup paired Bayesian optimization with optical evaluations of the Urbach energy, a metric used to evaluate sub-bandgap disorder. During heteroepitaxial growth runs, the autonomous system identified deposition conditions that reduced the Urbach energy to 182 meV, dropping below previously reported values for sputtered β-Ga2O3 films. The researchers found that the optimized growth window transferred across substrates, producing single-crystalline β-Ga2O3 homoepitaxy verified by scanning transmission electron microscopy.
To convert the closed-loop optimization data into interpretable growth rules, the team trained a random forest surrogate model on the experimental results. The model revealed that the growth landscape consists of additive contributions from four growth parameters alongside an interaction between temperature and oxygen.
That structure enabled a human-executable strategy featuring sequential one-dimensional tuning followed by focused two-dimensional refinement. Human-executed re-optimization using those derived rules lowered the Urbach energy further to 163 meV. The study was conducted by Yuki K. Wakabayashi, Yui Ogawa, Franz Benedict Romero, Coleman Wagner, and Yoshitaka Taniyasu, who reported no competing interests.
