Researchers at Japan's National Institute for Materials Science (NIMS) have developed a new system called pinax that captures the entire process of developing new materials, including machine learning workflows and decision-making processes. Published in the journal Science and Technology of Advanced Materials: Methods, the system addresses a critical challenge in materials science: the need to track not only experimental results but also the chain of reasoning behind them.
Machine learning models are increasingly used to discover and characterize new materials, but their reasoning processes are often opaque. Scientists may not know what trial-and-error steps led to final predictions. According to lead author Satoshi Minamoto of NIMS, "The system introduced in this study visualizes these invisible processes. This enables others to review, verify, and build upon the path to the conclusions." By formalizing both successful and unsuccessful trial-and-error processes, pinax enhances reproducibility, accountability, and knowledge sharing while maintaining strict data governance.
The importance of such transparency is underscored in applications where safety and accountability are paramount. Minamoto stated that this work "demonstrates how transparent AI systems can transform scientific discovery into a more reliable, efficient, and socially responsible endeavor." The system was tested using two case studies: predicting steel properties and using transfer learning to predict the thermal conductivity of polymers. In both cases, pinax made it possible to link model performance predictions to specific data or model aspects, and to reproduce complex multi-stage workflows. "In particular, the transfer-learning example highlights pinax's ability to track how information flows between intertwined datasets and models, making every step in the reasoning process explicitly traceable," Minamoto added.
The NIMS team plans to expand pinax towards an autonomous, closed-loop materials discovery system. By integrating pinax's tracking capabilities with automated experimental and simulation systems, they aim to create a loop that can use data generation, machine learning models, and decision-making systems to systematically and independently carry out the entire research cycle. The full details of the study are available in the paper at https://doi.org/10.1080/27660400.2026.2629051. Further information about the journal can be found at https://www.tandfonline.com/STAM-M.
