- Address
- 305-0044 茨城県つくば市並木1-1 [アクセス]
研究内容
- Keywords
Materials informatics, AI, Machine learning
I am a researcher specializing in materials informatics, artificial intelligence, and data-driven materials science. My work applies machine learning, natural language processing, and knowledge representation to extract insights from scientific literature, organize materials knowledge, and support materials discovery. I also work on inverse design problems, using autoencoder-based representation learning to construct compact materials design spaces and Monte Carlo Tree Search (MCTS) to efficiently explore these spaces and identify promising material candidates. My research further includes knowledge graphs, scientific text mining, literature clustering, and machine-learning approaches for materials data, with particular interest in battery materials and large-scale scientific knowledge analysis.
所属学会
JSAP
マテリアル基盤研究センター
Data-driven algorithms for materials innovation
Materials informatics, Machine learning, Natural Language Processing
概要
Development of machine learning algorithms for materials design and discovery.
Automatic construction of materials databases from related research papers using natural language processing.
Development of profiling serivce for mateials science reserchers.
Prediction of materials properties using regression methods.
新規性・独創性
Development of MDTS (Materials Design by Tree search) Python package for materials space exploration based on Monte Carlo tree search (MCTS) (https://github.com/tsudalab/MDTS).
Development of visualization service for NIMS SAMURAI researchers catalouge to capture research output and connect researchers with similar reserach interest using natural language processing.
内容
Materials design and discovery can be represented as selecting the optimal structure from a space of candidates that optimizes a target property. The selection is an iterative process, where selected candidate in one iteration is evaluted and fedback to the process for a more informed selection in the next iteration. Since the number of candidates can be exponentially proportional to the structure determination variables, the efficiency to obtain the optimal structure is a critical issue. We use Monte Carlo tree search (MCTS) approach in combination with expansion policy neural network to accelerate this process. MCTS has no tuning parameters and works autonomously in various problems. Additionally, it does not require pre- availalbe training data and it is highly scalble making it good fit for large spaces problems such as the chemical and mateirals space.
まとめ
Optimization of the chemical compositions and heat treatment scheduling for new Ni-based superalloy for addiditve manufacturing using machine learning. Active-Leaning-Driven Development of Platinum-Free Electrocatalysts for the Hydrogen Oxidation Reaction.
