Bayesian Optimization with Experimental Failure for High-Throughput Materials Growth

Recent developments in materials informatics leveraging machine learning methods, including Bayesian optimization (BO) and artificial neural networks, have been accelerating materials research [1,2]. However, appropriate handling of missing data caused by experimental failures has been a common but significant challenge for the application of machine learning and automation technologies, such as BO and robotic experimentation, toward efficient high-throughput materials synthesis.<br/>BO is a sample-efficient approach for global optimization. In this presentation, we will introduce a novel BO algorithm that effectively handles missing data. Missing data instances arise when the target material is evasive due to inadequate synthesis parameters [3]. Our algorithm aims to complement and address this issue, enabling more robust and efficient optimization even in the presence of missing data. Though one may restrict the parameter search space to avoid experimental failures, this approach can exclude good parameters for the target material. In contrast, our algorithm can fully search complex parameter spaces by substituting the evaluation value for the missing data to the worst evaluation value available at that time. First, we show the efficacy of the BO method in handling experimental failures by utilizing simulated data. Subsequently, we validate its performance by implementing real materials growth data, namely machine-learning-assisted molecular beam epitaxy (ML-MBE) [4,5] of itinerant ferromagnetic perovskite SrRuO₃ thin films. We used the residual resistivity ratio (RRR) as the evaluation metric. The growth conditions, specifically the Ru/Sr flux ratio, growth temperature, and ozone flux rate, of SrRuO₃ thin films on DyScO₃ substrates were optimized for achieving high RRR values. The optimization process uses five random initial growth parameters and measured experimental RRR values for the updated Gaussian process regression model that predicted RRR values at unseen growth parameters using the past observations. We achieved the RRR of 80.1, the highest ever reported among tensile-strained SrRuO₃ films. We explored a wide three-dimensional parameter space, while complementing missing data within only 35 MBE growth runs. Our tensile-strained SrRuO₃ thin films stabilized by epitaxial strain show higher Curie temperature than bulk or compressive-strained films [6]. The proposed BO method is capable to properly handle experimental failure and will play an essential role in the growth/synthesis of various materials.<br/><br/><b>References</b> [1] Mueller <i>et al</i>., Reviews in Computational Chemistry <b>29</b> (Wiley, Hoboken, 2015). [2] F. Ren <i>et al., </i>Sci. Adv. <b>4</b>, eaaq1556 (2018). [3] Y. K. Wakabayashi^*, T. Otsuka^*<i> et al</i>., npj Comput. Mater. <b>8</b>, 180 (2022) [4] Y. K. Wakabayashi^*, <i>et al</i>., APL Mater. <b>7</b>, 101114 (2019). [5] K. Takiguchi, Y. K. Wakabayashi^*,<i> et al</i>., Nat. Commun. <b>11</b>, 4969 (2020). [6] G. Koster <i>et al., </i>Rev. Mod. Phys. <b>84</b>, 253 (2012).