Multicrystalline informatics: a methodology to advance materials science by unraveling complex phenomena.

Multicrystalline materials play a crucial role in our society. However, their microstructure is complicated, and there is no universal approach to achieving high performance. Therefore, a methodology is necessary to answer the fundamental question of how we should design and create microstructures. 'Multicrystalline informatics' is an innovative approach that combines experimental, theoretical, computational, and data sciences. This approach helps us understand complex phenomena in multicrystalline materials and improve their performance. The paper covers various original research bases of multicrystalline informatics, such as the three-dimensional visualization of crystal defects in multicrystalline materials, the machine learning model for predicting crystal orientation distribution, network analysis of multicrystalline structures, computational methods using artificial neural network interatomic potentials, and so on. The integration of these research bases proves to be useful in understanding unexplained phenomena in complex multicrystalline materials. The paper also presents examples of efficient optimization of the growth process of high-quality materials with the aid of informatics, as well as prospects for extending the methodology to other materials.

We introduce our attempt to explore a methodology for high-performance multicrystalline materials development, multicrystalline informatics, using silicon as a model material.

Bayesian Optimization for Cascade-Type Multistage Processes.

Complex processes in science and engineering are often formulated as multistage decision-making problems. In this letter, we consider a cascade process, a type of multistage decision-making process. This is a multistage process in which the output of one stage is used as an input for the subsequent stage. When the cost of each stage is expensive, it is difficult to search for the optimal controllable parameters for each stage exhaustively. To address this problem, we formulate the optimization of the cascade process as an extension of the Bayesian optimization framework and propose two types of acquisition functions based on credible intervals and expected improvement. We investigate the theoretical properties of the proposed acquisition functions and demonstrate their effectiveness through numerical experiments. In addition, we consider suspension setting, an extension in which we are allowed to suspend the cascade process at the middle of the multistage decision-making process that often arises in practical problems. We apply the proposed method in a test problem involving a solar cell simulator, the motivation for this study.

High Passivation Performance of Cat-CVD i-a-Si:H Derived from Bayesian Optimization with Practical Constraints.

High-quality passivation with intrinsic hydrogenated amorphous Si (i-a-Si:H) is essential for achieving high-efficiency Si heterojunction (SHJ) solar cells. The formation of i-a-Si:H with a high passivation quality requires strict control of the hydrogen content and film density. In this study, we report the effective discovery of i-a-Si:H deposition conditions through catalytic chemical vapor deposition using Bayesian optimization (BO) to maximize the passivation performance. Another contribution of this study to materials science is the establishment of a practical BO scheme consisting of several prediction models in order to account for the practical constraints. By applying the BO scheme, effective minority carrier lifetime (τeff) is maximized within the deposition condition range, while being constrained by the i-a-Si:H thickness and the capabilities of the experimental setup. We achieved a high passivation performance of τeff > 2.6 ms with only 8 cycles in BO, starting with 14 initial samples. Within the investigated range, the deposition conditions were further explored over 20 cycles. The BO provided not only optimal deposition conditions but also scientific knowledge. Contour plots of the predicted τeff values obtained through the BO process demonstrated that there is a band-like high τeff condition in the parameter space between the substrate temperature and SiH4 flow rate. The high void fraction and epitaxial growth were inhibited by controlling the substrate temperature and SiH4 flow rate, resulting in a high passivation quality. This indicates that the combination of the SiH4 flow rate and substrate temperature parameters is crucial to passivation quality. These results can be applied to determine the deposition conditions for a good a-Si:H layer without a high void fraction or epitaxial growth. The research methods shown in this study, practical BO scheme, and further analysis based on the optimized results will be also useful to optimize and analyze the process conditions of semiconductor processes including plasma-enhanced chemical vapor deposition for SHJ solar cells.

Multicrystalline Informatics Applied to Multicrystalline Silicon for Unraveling the Microscopic Root Cause of Dislocation Generation.

A comprehensive analysis of optical and photoluminescence images obtained from practical multicrystalline silicon wafers is conducted, utilizing various machine learning models for dislocation cluster region extraction, grain segmentation, and crystal orientation prediction. As a result, a realistic 3D model that includes the generation point of dislocation clusters is built. Finite element stress analysis on the 3D model coupled with crystal growth simulation reveals inhomogeneous and complex stress distribution and that dislocation clusters are frequently formed along the slip plane with the highest shear stress among twelve equivalents, concentrated along bending grain boundaries (GBs). Multiscale analysis of the extracted GBs near the generation point of dislocation clusters combined with ab initio calculations has shown that the dislocation generation due to the concentration of shear stress is caused by the nanofacet formation associated with GB bending. This mechanism cannot be captured by the Haasen-Alexander-Sumino model. Thus, this research method reveals the existence of a dislocation generation mechanism unique to the multicrystalline structure. Multicrystalline informatics linking experimental, theoretical, computational, and data science on multicrystalline materials at multiple scales is expected to contribute to the advancement of materials science by unraveling complex phenomena in various multicrystalline materials.

Bayesian optimization of hydrogen plasma treatment in silicon quantum dot multilayer and application to solar cells.

Silicon quantum dot multilayer (Si-QDML) is a promising material for a light absorber of all silicon tandem solar cells due to tunable bandgap energy in a wide range depending on the silicon quantum dot (Si-QD) size, which is possible to overcome the Shockley-Queisser limit. Since solar cell performance is degenerated by carrier recombination through dangling bonds (DBs) in Si-QDML, hydrogen termination of DBs is crucial. Hydrogen plasma treatment (HPT) is one of the methods to introduce hydrogen into Si-QDML. However, HPT has a large number of process parameters. In this study, we employed Bayesian optimization (BO) for the efficient survey of HPT process parameters. Photosensitivity (PS) was adopted as the indicator to be maximized in BO. PS (σp/σd) was calculated as the ratio of photoconductivity (σp) and dark conductivity (σd) of Si-QDML, which allowed the evaluation of important electrical characteristics in solar cells easily without fabricating process-intensive devices. 40-period layers for Si-QDML were prepared by plasma-enhanced chemical vapor deposition method and post-annealing onto quartz substrates. Ten samples were prepared by HPT under random conditions as initial data for BO. By repeating calculations and experiments, the PS was successfully improved from 22.7 to 347.2 with a small number of experiments. In addition, Si-QD solar cells were fabricated with optimized HPT process parameters; open-circuit voltage (VOC) and fill factor (FF) values of 689 mV and 0.67, respectively, were achieved. These values are the highest for this type of device, which were achieved through an unprecedented attempt to combine HPT and BO. These results prove that BO is effective in accelerating the optimization of practical process parameters in a multidimensional parameter space, even for novel indicators such as PS.

Data-Driven Optimization and Experimental Validation for the Lab-Scale Mono-Like Silicon Ingot Growth by Directional Solidification.

The casting mono-like silicon (Si) grown by directional solidification (DS) is promising for high-efficiency solar cells. However, high dislocation clusters around the top region are still the practical drawbacks, which limit its competitiveness to the monocrystalline Si. To optimize the DS-Si process, we applied the framework, which integrates the growing experiments, transient global simulations, artificial neuron network (ANN) training, and genetic algorithms (GAs). First, we grew the Si ingot by the original recipe and reproduced it with transient global modeling. Second, predictions of the Si ingot domain from different recipes were used to train the ANN, which acts as the instant predictor of ingot properties from specific recipes. Finally, the GA equipped with the predictor searched for the optimal recipe according to multi-objective combination, such as the lowest residual stress and dislocation density. We also implemented the optimal recipe in our mono-like DS-Si process for verification and comparison. According to the optimal recipe, we could reduce the dislocation density and smooth the growth rate during the Si ingot growing process. Comparisons of the growth interface and grain boundary evolutions showed the decrease of the interface concavity and the multi-crystallization in the top part of the ingot. The well-trained ANN combined with the GA could derive the optimal growth parameter combinations instantly and quantitatively for the multi-objective processes.