NTT ускорила тестирование пленок в три раза с помощью новой платформы на базе искусственного интеллекта

NTT triples film-testing speed with beta-gallium oxide AI platform KEY POINTS - NTT demonstrates autonomous film deposition system that runs experiment cycles about three times faster - Platform produces single-crystal beta-gallium oxide thin film by sputtering and extracts human-usable deposition rules - Nature Communications publishes the research, highlighting NTT's AI-for-science push in semiconductor materials NTT said on August 25 it had demonstrated an "interpretable autonomous deposition" system that uses artificial intelligence and automation to optimize thin-film growth while converting experimental data into deposition rules that researchers can understand and reuse. The Japanese telecom and technology company said the platform links automated sputter deposition, automated optical evaluation and machine-learning-based selection of the next process conditions in a closed loop. The system repeats deposition, evaluation and condition-setting without human judgment or analysis between steps, enabling experiment cycles to run about three times faster than conventional operation by engineers and researchers. Using the platform, NTT produced a high-quality single-crystal thin film of ultrawide-bandgap semiconductor beta-gallium oxide, or β-Ga2O3, by sputtering, which it described as a world first for that deposition method. Beta-gallium oxide is being studied as a next-generation power semiconductor material. The autonomous optimization targeted four deposition parameters: temperature, sputtering power, argon flow and oxygen flow. NTT applied the system to β-Ga2O3 film formation on sapphire wafers, using Urbach energy, a measure tied to film quality, as the optimization index. After 56 autonomous search cycles, it identified conditions that produced an Urbach energy of 182 meV, which NTT said was the lowest reported for sputtered β-Ga2O3 films formed on sapphire wafers. NTT then transferred the resulting conditions to deposition on a β-Ga2O3 substrate and achieved a high-quality single-crystal thin film. The sapphire-based films also showed heteroepitaxial growth composed only of the target β-Ga2O3 crystal phase, without different crystal phases. Beyond finding optimal conditions, NTT analyzed data from the autonomous feedback loop using a random forest model to break down the contributions of each deposition parameter and their interactions. Most changes in film quality could be predicted by adding the effects of temperature, sputtering power, argon flow and oxygen flow, while the combination of temperature and oxygen flow stood out as an interaction requiring closer adjustment. The company reorganized the model's predictions into a simpler form that researchers could interpret, consisting of each parameter's contribution and the temperature-oxygen flow interaction. Researchers then used those extracted rules for re-optimization, reducing Urbach energy further to 163 meV and confirming that the rules were effective as a practical guide for improving deposition conditions. The research was published in Nature Communications on August 13 under the title "Interpretable self-driving sputtering epitaxy reveals human-usable growth rules for β-Ga2O3 films." AI for Science has drawn growing attention in materials research as laboratories seek to automate experimental workflows and use machine learning to navigate large parameter sets more efficiently. In semiconductor deposition, where film quality depends on multiple process variables such as temperature and gas flow, faster optimization can shorten development cycles for materials used in power devices. NTT said it plans to expand the approach to optimization using multiple evaluation metrics, enhance AI models with knowledge of material properties and deposition processes, and test the system across a wider range of materials and equipment. © Copyright 2026 IBTimes JP. All rights reserved.