Harnessing NOMAD as a Collaborative Platform for Scientific Machine Learning

As scientific machine learning (ML) becomes increasingly prevalent across disciplines, the need for FAIR (Findable, Accessible, Interoperable, Reusable) data practices and domain-specific contextualization is more critical than ever. NOMAD [1], originally developed to store and share materials science data, is evolving into a collaborative environment that supports the full lifecycle of scientific research in materials science, including machine learning-driven workflows from data ingestion to model sharing. This contribution explores recent extensions to NOMAD that enable researchers to integrate ML workflows with experimental and simulation data. Central to this capability is NOMAD's powerful data modeling approach: metadata schemas that capture the scientific semantics of the data, enabling precise search queries. Extending this, NOMAD supports the storage and indexing of ML models themselves, enriched with metadata that describes their training data and scientific context, along with the architectural and training details. Customizable search dashboards in NOMAD allow to search ML models based on the metadata intuitively. We showcase this functionality through the nomad-auto-xrd plugin [2], which enables the training and application of ML models for automated X-ray diffraction (XRD) analysis. Built on the XRD AutoAnalyser framework developed by Szymanski et al. [3], the plugin wraps this functionality to operate directly on NOMAD entries. Users can define new model entries by specifying the compositional space of interest, parameters for generating simulated training data, and details about the model architecture and training process. Based on these specifications, the training can be performed using NORTH, an integrated toolkit that allows users to run Jupyter notebooks directly within NOMAD. The trained models are indexed as NOMAD entries enriched with metadata describing their scientific context. This enables users to search for models tailored to specific chemical spaces or experimental conditions relevant to their samples. As a result, users can easily identify and reuse appropriate AutoXRD models, connect them to experimental XRD entries, and perform automatic phase identification within a unified infrastructure. The entire workflow—from data generation to model inference—is traceable, transparent, and reproducible, promoting reuse and collaborative refinement across the scientific community. NOMAD's integration of data and model provenance, advanced searchability, and standards-driven interoperability demonstrates a scalable approach to managing scientific ML assets linked with rich domain-specific knowledge in line with FAIR principles. [1] Scheidgen et al., (2023). NOMAD: A distributed web-based platform for managing materials science research data. Journal of Open Source Software, 8(90), 5388, https://doi.org/10.21105/joss.05388 [2] Marquez, P., & Kapoor, S. (2025). nomad-auto-xrd [Computer software]. https://github.com/FAIRmat-NFDI/nomad-auto-xrd [3] Szymanski, N.J., Fu, S., Persson, E. et al. Integrated analysis of X-ray diffraction patterns and pair distribution functions for machine-learned phase identification. npj Comput Mater 10, 45 (2024). https://doi.org/10.1038/s41524-024-01230-9