Machine Learning

W ith recent advances in machine learning technology, data-driven research is beginning to permeate natural science and engineering fields. Synchrotron radiation science is also expected to benefit significantly from machine learning. The progress of these studies will make it possible to observe materials that could not be observed in the past or to perform synchrotron radiation measurements and detailed data analysis much more efficiently than before, leading to more effective use of limited beamtime. In addition, machine learning has the potential to bring about advanced and more efficient research through software without the need for major hardware upgrades at synchrotron radiation facilities. The encounter between machine learning and materials science has opened up a new academic field called materials informatics. Especially in the last decades, the progress has been remarkable, and the concept of informatics has been incorporated into all areas of materials science, from material design and material synthesis to measurement and analysis. The rise of materials informatics was due to advances in information science in terms of both hardware and software; namely, the dramatic development of computing power and artificial intelligence technologies such as machine learning, which have made it possible to handle large volumes of complex data that were difficult to handle in the past. In addition, it is now possible to extract useful information and new knowledge from the data, bringing about changes in various fields. Furthermore, machine learning technology has become much easier than in the past, thanks not only to simple programming languages such as Python but also to open source platforms on which an ecosystem for data analysis has been built. Taking synchrotron radiation experiments as an example, the measurement space to be explored in experiments is extremely wide. In order to extract knowledge from complex data analysis, it is necessary to efficiently search a high-dimensional search space consisting of an enormous number of parameters to find the optimal solution. Parameter search in such a highdimensional space, which skilled experts conventionally conduct based on tacit knowledge such as intuition and experience, poses problems such as bottlenecks to automation, human bias, and poor reproducibility, and requires a new research methodology that will fundamentally change conventional research methods. The wide range of new developments in the combination of synchrotron radiation and machine learning discussed in this special issue will extend synchrotron radiation experiments to more advanced measurements, bring about more efficient and automated synchrotron radiation experiments, and increase the amount of information obtained from these experiments. We hope these efforts will contribute significantly to further developing and revitalizing the synchrotron radiation community and opening up new research fields. n Kanta Ono Guest Editor Osaka University, Osaka, Japan ono@ap.eng.osaka-u.ac.jp Synchrotron Radiation News ISSN 0894-0886 is published bi-monthly. Coden Code: SRN EFR