We develop a method for calculating the minimum X-ray exposure time for X-ray solution scattering experiments using statistical and mathematical approaches to enhance the measurement efficiency while maintaining data quality. Experts can determine the X-ray exposure time for samples by investigating time-series data of scattering profiles and considering the quality of scattering profiles and the samples’ X-ray damage. In this study, we apply a statistical inequality to estimate the kernel density estimation method’s error to determine the minimum X-ray exposure time for protein solution scattering experiments and evaluate the method’s validity using the results of Guinier analysis as an indicator.

Building machine learning models using a dataset calculated by first principles calculations is an important approach to explore the next-generation batteries. In previous studies, the rechargeable secondary battery dataset was constructed for Li ion by density functional theory (DFT) calculation, and after that, extended to five alkali metal ions. This dataset can be regarded as consisting of five alkali metal ion groups, and it is one of the interests to know which approach is preferred to build individual models specialized for each alkali metal ion or build a single model by integrating the datasets. We quantitatively evaluate the possibility of data integration in the framework of Bayesian model selection and show that the integration of datasets is suitable. In addition, extracting new knowledge using feature selection is also important in exploring next-generation batteries. In order to further advance the knowledge extraction, the reliability of the selected features should be considered to avoid misinterpretation. We evaluate the confidence level of feature selection by calculating the posterior probabilities of features using Bayesian model averaging (BMA). We found that the confidence level of feature selection increases when the number of data increases.

A virtual-center-atom (VCA) structural descriptor was proposed to construct machine learning potentials which can be applied in studying alloy materials. The descriptor performs by describing the spatial features of atomic environments centered on a defined virtual atom within crystalline structure. Computational algorithms were developed to determine the position of virtual center atom. To verify the generality of this descriptor, crystal structures from both molecular dynamics trajectory and online accessible databases are utilized to check their machine learning (ML) performances. The challenge of the ML lies in increasing the transferability of the derived weight parameters to cover a broad range of structure distribution. Forces are found difficult to train the ML model under simple optimization methods. We thus suggest deep learning techniques for improving the applicability of the VCA potential. The descriptor is simple compared to other reported schemes and has promising applications in accelerating material discovery.

The new fatigue data sheet, No. 132, discloses gigacycle fatigue properties of the A6061-T6 aluminium alloys at high stress ratios. The fatigue tests were conducted mainly by the ultrasonic fatigue testing at 20 kHz, while conventional fatigue tests at 100 Hz were also conducted for comparison. The fatigue test results indicated that fatigue limits were obscure in the A6061-T6 alloys. Many specimens failed at over 107 cycles, developing internal fractures as well as surface fractures. On the other hand, fatigue failures at over 109 cycles were very rare, suggesting the presence of new fatigue limits in the gigacycle region. The differences in fatigue test results were negligible between the 20 kHz and 100 Hz tests, demonstrating that the 20 kHz tests were comparable to the conventional fatigue tests on this material. The fatigue strengths evaluated by the stress amplitude were decreased according to increase in the stress ratios, while the degradation of the fatigue strengths was not so serious. The fatigue strengths were higher than the modified Goodman lines, meaning that the stress ratio effects could be estimated by conventional ways.

To overcome the trade-off relationship between tensile strength and elongation of dual-phase steels, three exploratory techniques were utilized: Bayesian optimization (BO), BoundLess Objective-free eXploration (BLOX), and one-class support vector machine (OCSVM). The BO is an optimization method using Gaussian process regression, and the BLOX and OCSVM are designed for finding out-of-trend materials. Initially, a large number of synthetic phase distributions of ferrite and martensite were prepared by the Gaussian random field method. Feature importance analysis was then performed to extract microstructural descriptors used as explanatory variables in the explorations. The results revealed that the volume fraction of martensite, the higher-order principal component of the spatial correlation function, and the first principal component of the persistent homology are important in predicting the product of strength and elongation (TS × EL). In the three exploration methods, crystal plasticity finite element analyses with a ductile damage model were repeatedly performed to predict strength and elongation. The BO and OCSVM successfully found microstructures with TS × EL much higher than those of random search in a limited number of iterations. The microstructures discovered in the explorations indicated that two approaches to material design are effective in improving tensile properties of dual-phase steels: uniformly dispersing the fine hard phase and constructing a lamellar structure of the hard and soft phases.

Here, we propose an interaction analysis method, which quantitatively evaluates the interactions among constituent materials in composites that consist of matrix (e.g. resin), fillers, and additives (e.g. fiber-reinforced plastics). Using this method, we developed a data-driven scheme that is capable of identifying alternative constituent materials. The scheme enabled the search for new composites with the same performance as the original composite. The proposed scheme was also found useful for composites with complex interactions, where the features of constituent materials (e.g. the addition ratio) exhibited linear and exponential relationships with the performance. To demonstrate the utility of the proposed scheme, we searched for alternative constituent materials, which are involved in the interaction with the largest contribution to the performance (i.e. bending modulus). The bending moduli of the resulting composites were comparable to the original composite, thereby verifying the utility of the proposed scheme.

Materials exploration requires the optimization of a multidimensional space including the chemical composition and synthesis parameters such as temperature and pressure. Bayesian optimization has attracted attention as a method for efficient multidimensional optimization. Appropriate choices of the acquisition function and initial values of the hyperparameters of the kernel functions are essential for the Bayesian optimization of synthesis conditions in a small number of experiments. However, to date, there has been little discussion on how to tune Bayesian optimization for materials exploration, and no guidelines have been provided for materials scientists. In this study, we investigated the optimum initial values of the hyperparameters in Bayesian optimization using one-dimensional model functions that mimic actual materials syntheses. The optimal lengthscale and variance for different process windows of materials synthesis were investigated. It was shown that the use of an appropriate acquisition function and suitable initial values of the hyperparameters of the kernel functions enable the optimization of synthesis conditions in a small number of trials. These results provide insight for enabling fully automated and autonomous materials synthesis using Bayesian optimization and robotics for materials exploration.

We propose a highly efficient searching algorithm in crystal structure prediction. The searching algorithm is a hybrid of the evolutionary algorithm and Bayesian optimization. The evolutionary algorithm is widely used in crystal structure prediction, and the Bayesian optimization is one of the selection-type algorithms we have developed. We have performed simulations of crystal structure prediction to compare the success rates of the random search, evolutionary algorithm, Bayesian optimization, and hybrid algorithm for up to ternary systems such as Si, Y2Co17, Al2O3, and CuGaS2, using the CrySPY code. These results demonstrate that the evolutionary algorithm can generate structures more efficiently than random structure generation, and the Bayesian optimization can efficiently select potential candidates from a large number of structures. Moreover, the hybrid algorithm, which has the advantages of both, is proved to be the most efficient searching algorithm among them.