NIMS Engineers Develop System to Track Material Design Processes, Enhancing AI Transparency and Reproducibility
Engineers at Japan's National Institute for Materials Science have developed a system called pinax that captures the entire trial-and-error process in material design, addressing critical challenges in reproducibility and accountability for AI-driven scientific discovery.
Engineers at Japan's National Institute for Materials Science (NIMS) have developed a system that captures all elements of trial and error in material design, enabling reliable reproduction of reasoning processes and results. The system, called pinax, addresses significant challenges in materials discovery where machine learning models play an increasingly important role but often operate as opaque tools whose decision-making processes remain hidden.
Published in the journal Science and Technology of Advanced Materials: Methods, pinax formalizes both successful and unsuccessful trial-and-error processes to enhance reproducibility, accountability, and knowledge sharing while maintaining strict data governance. According to lead author Satoshi Minamoto of NIMS, this approach transforms scientific discovery into a more reliable, efficient, and socially responsible endeavor by making invisible processes visible and reviewable.
The system's development responds to growing needs in fields like clean energy, advanced manufacturing, and improved infrastructure, where discovering and characterizing new materials drives technological progress. Researchers generate large amounts of experimental and computational data but have lacked tools to track and store not only results but also the chain of reasoning behind them. Pinax captures the entire process of developing new materials, including machine learning workflows and decision-making processes that typically remain undocumented.
Minamoto emphasizes the importance of such transparency in applications where safety, reproducibility, and accountability are critical. The system enables others to review, verify, and build upon the path to conclusions, addressing concerns about the opaque nature of many machine learning models used in materials discovery. Researchers tested pinax using two case studies: one predicting steel properties and another using transfer learning to predict polymer thermal conductivity. The system successfully linked model performance predictions to specific data or model aspects that influenced them and reproduced complex, multi-stage workflows.
Particularly in the transfer-learning example, pinax demonstrated its ability to track how information flows between intertwined datasets and models, making every step in the reasoning process explicitly traceable. The engineers now plan to expand pinax toward an autonomous, closed-loop materials discovery system by integrating its tracking capabilities with automated experimental and simulation systems. This integration aims to create a loop that can systematically and independently carry out the entire research cycle using data generation, machine learning models, and decision-making systems. The research paper detailing this development is available at https://doi.org/10.1080/27660400.2026.2629051, and further information about the journal can be found at https://www.tandfonline.com/STAM-M.