Data-Driven Researches Under the Concept of Materials Integration

The concept of Materials Integration is to computationally link among processing, structure, property, and performance for accelerating materials research and development [1,2] and is comparable to that of Integrated Computational Materials Engineering. To embody the Materials Integration, we have developed a system named MInt (Materials Integration by Network Technology) [3]. In MInt, one can implement any types of prediction models based on computational simulations, theoretical or empirical formulas, and machine learning algorithms as a module. Furthermore, one can connect several modules to design a workflow that comprehensively predicts from processing through microstructure to property and/or performance. MInt executes a workflow by delivering the output data from a module to the input port of the next module and stores all the computational results in order. MInt is a platform to accumulate the materials scientists’ knowledge digitalized as modules and workflows, which are stored not in isolation but in the form of graph network. At this moment, the modules and workflows implemented in MInt are mainly for conventional structural alloys such as steels, aluminum, and Ni-based superalloys, but one can expand its application fields by adding new modules and workflows.<br/><br/>Data-driven approach plays an important role in Materials Integration. Firstly, it is useful to make the prediction modules more accurate; for example, data assimilation technique can adjust model parameters included in the computational simulation or the theoretical or empirical formula to improve predictability. Secondary, it helps to build a comprehensive workflow even if all the mechanisms are not fully understood. For a missing link, machine learning provides a non-linear regression model based on experimental data without physical background. On top of that, certain data-driven methods, such as sparse modeling and feature importance analysis, contribute to finding the physical picture. Lastly, it provides a realistic means of solving the inverse problem, i.e., finding the optimal material and processing that meets the desired performance. Since MInt provides an application programing interface to execute the workflow and extract the computed results, one can apply any types of optimization algorithms including Bayesian optimization, Monte-Carlo tress search and so on. In this talk, I would like to share some examples of where the data-driven approach has worked.<br/><br/>[1] Demura M, Koseki T. Mater Japan. 2019;58(9):489-493. doi:10.2320/materia.58.489 [2] Demura M. Mater Trans. 2021;62(11):1669-1672. doi:10.2320/matertrans.MT-M2021135 [3] Minamoto S, Kadohira T, Ito K, Watanabe M. Mater Trans. 2020;61(11):2067-2071. doi:10.2320/matertrans.MT-MA2020002