Data analysis and machine learning are being used in the research, development and production of functional materials such as crystal materials, enabling more efficient design of molecules, materials and processes and process control. After constructing a nonlinear machine learning model, new molecules, materials and processes can be proposed with direct inverse analysis of the model to perform molecules, materials and processes directly from target values of properties and activities. In addition, the model can be interpreted using genetic algorithm-based partial least squares regression with only the first component, cross-validated permutation feature importance, and local slope of model prediction to discuss important parameters in experiments, manufacturing, and processes, and to advance understanding of phenomena and clarification of mechanisms in molecules and materials.

  Bayesian optimization has been attracting attention as an efficient method for synthesizing new materials. While Bayesian optimization is a powerful technique, applying it to material synthesis requires appropriate goal setting and tuning based on the experience and knowledge of materials researchers. Therefore, this manuscript provides an overview of Bayesian optimization and important considerations for materials researchers.

  In recent years, the development of metal−organic framework (MOF) structural databases for computational science and machine learning has been actively pursued, with these databases being used to explore adsorbents for gas storage and separation. However, no reports have focused on screenings that highlight the adsorption-induced structural transitions unique to MOFs having flexible framework (soft MOFs). This gap is due to the absence of established computational methodologies. In this explanatory article, we discuss the computational screening of soft MOFs using the thermodynamic integration method.

  A crystal structure search is progressing, utilizing a genetic algorithm that heuristically updates the crystal structure by applying genetic evolution processes such as crossover and mutation to its unit cell and constituent atoms therein. In this article, we present our recent achievement in applying this approach to the search for new ternary hydrides that exhibit high superconducting transition temperatures. Intrinsically, the heuristic search only yields results like, "we don’t know why, but it works." To understand such discoveries, we can utilize first-principles calculations to hypothesize why the desired properties are achieved. Incorporating this knowledge allows us to impose constraints on the excessively broad search space during the heuristic updates, thereby making material exploration more efficient. This article highlights a case where such a scheme has been highly successful. Heuristic approaches combined with machine learning would accelerate the exploration of unknown crystal structures that exhibit desired physical properties.

  Deep learning has been gaining attention in various fields over this decade. This class of techniques succeeded in many areas, e.g., image recognition and natural language processing. However, deep learning has yet to be fully applied in some areas due to reasons such as a lack of large-scale training datasets. This paper briefly introduces deep learning and some research examples of deep learning in which the author was involved. The first example demonstrated that countless training data were able to be generated automatically by designing a model and a loss function, and the output behaviors could be fine-tuned by adjusting internal feature variables. The second example demonstrated that the generalization ability of a motion-generation model could be remarkably improved by applying neuroscientific findings to the model structure. Finally, this paper introduces force control as an essential concept in robot applications, including laboratory automation.