Machine-Learning Minimization of Amorphous Heat Conduction by High-Throughput Deposition Process in The Loop

In Material Informatics, material exploration <i>via</i> high-throughput experiments has gained traction. This study utilizes a specially equipped combinatorial sputtering system and thermoreflectance method to explore amorphous materials with TC ~0.1 W/mK. There is a significant demand for solid-like thermal insulation materials combining mechanical strength, which can be applied to heat-mediated sensing devices and heat shields. Amorphous materials are promising candidates for dense thermal insulator since the disordered structure strongly suppress propagation of the vibrational modes. Their metastable nature can result in varied structures depending on the fabrication process. For instance, our prior research demonstrated that TC of amorphous Si (a-Si) and Ge (a-Ge) films varies with the deposition temperature during sputtering. This implies that amorphous structures can be modulated using sputtering process parameters (PP). Due to challenges in identifying structures, optimizing properties of amorphous materials has not been studied much. In this study, while treating the structure and the composition as a black box, we use the sputtering PPs as the explanatory variables and TC as the target variable to be minimized through Bayesian optimization. This high-throughput experiment also works as a materials screening process to identify interesting materials for the characterization analysis. It allows efficient experimental data accumulation on the relationships between the thermal conductivity and the characteristics of amorphous. As a further objective of this study, we also plan to implement machine learning to interpret the gathered data and enhance our comprehension of heat transfer in amorphous materials, which is not yet fully understood.<br/>Samples were prepared using Combinatorial Sputter Coating System (COSCOS), which can deposit the target material with four arbitrarily controlled PPs: mixed gas composition, total gas pressure, sputtering power, and sample-target distance. The multiple holders enable automated deposition of 14 samples under a unique set of PPs, in a single batch. We prepared 40 of a-SiO<i>_x</i>, 35 of a-SiN<i>_x</i>, and 12 of a-Si<i>_a</i>Ge<i>_b</i>Sn<i>_c</i> films with a thickness between 30-200 nm and measured the TC. Here, a-SiO<i>_x</i> and a-SiN<i>_x</i> were measured by frequency-domain thermoreflectance (FDTR) and a-Si<i>_a</i>Ge<i>_b</i>Sn<i>_c</i> was measured by time-domain thermoreflectance (TDTR). Moreover, for the a-SiN<i>_x</i>, deposition was performed under 26 new conditions, as suggested by Bayesian optimization.<br/>32 out of the 40 a-SiO<i>_x</i> films exhibited similar TC ranging from 0.6 to 0.9 W/mK except for the remaining 8 films with TC between 1.4-2.2 W/mK, likely due to contaminations of iron clusters from the equipment during sputtering. In contrast, a-SiN<i>_x</i> exhibited TC ranges from 0.6 to 1.5 W/mK with a wide-tailed distribution. This indicates the broader TC modulation range for a-SiN<i>_x</i>, compared with a-SiO<i>_x</i>. Among 26 of a-SiN<i>_x</i> films prepared under PPs suggested by Bayesian optimization, the one with the lowest TC achieved 0.5 W/mK. This result ensures the efficacy of our strategy to reduce TC. For a-Si<i>_a</i>Ge<i>_b</i>Sn<i>_c</i> films, the lowest TC was 0.2 W/mK, lower than the reported TC of similar materials like a-Ge₃₉Si₅Sn_1.3O₃₆ (0.4 W/mK). TDTR also confirmed decreases in speed of sound as Sn content increases which contributes to the TC reduction. As a result, a-Si<i>_a</i>Ge<i>_b</i>Sn<i>_c</i> emerges as a promising candidate for further optimization. Various characteristics like density and structure, beyond merely composition, might be modulated through the depositions across diverse PPs.<br/>To enhance the experimental throughput, we employed the masking mechanism of COSCOS for multi-point deposition on a single substrate and FDTR/TDTR mapping measurements. This experimental scheme should facilitate data accumulation on the order of 100 samples per single deposition batch. We are currently continuing our material exploration in the a-Si<i>_a</i>Ge<i>_b</i>Sn<i>_c</i> system and will report on these results in our presentation.