Cross-Facility Materials Discovery Platform: Integrating Heterogeneous Data and Automated Experimental Pipelines into High-Throughput Workflows
Artificial intelligence (AI)-driven materials discovery depends on the availability of large, highly annotated experimental datasets. However, data generated in typical materials science projects are fragmented across heterogeneous instruments, laboratories, and user facilities, each with its own formats, schedules, and degree of curation. In addition, manual sample preparation, time-consuming characterization, and changing sample form factors and holders between process steps often limit throughput and therefore dataset size. Here, we present a modular cross-facility workflow that addresses both the physical and the data bottlenecks. We combine robotic thin-film deposition, a custom robotic UV-vis setup, and high-throughput synchrotron end stations with custom, QR-code-tagged sample holders, preserving each film’s identity as it moves between experimental setups. Using Crucible, a schema-free data management platform developed at Lawrence Berkeley National Laboratory, we link every sample, container, and dataset into a provenance network that traces each thin film through synthesis, deposition, characterization, and analysis. A Python client and custom web interfaces connect the Molecular Foundry with two beamlines at the Advanced Light Source and enable streamlined, high-throughput data collection pipelines. To demonstrate this workflow, we fabricated 6,105 thin films spanning 618 hybrid metal halide compounds over six months, of which approximately 3,400 were characterized by robotic grazing-incidence wide-angle X-ray scattering and 2,000 by residual gas analysis under extreme-ultraviolet exposure, generating a highly curated AI-ready materials discovery dataset that accurately reproduces a subset of matched manually prepared samples. We further demonstrate automated downstream analysis, including human-in-the-loop machine-learning classification and mechanistic evaluation of structure-property relationships. Together, the dataset and infrastructure provide a blueprint for scaling materials experimentation to facility-level AI platforms.
Authors
- Wiebke Koepp (ORCID: https://orcid.org/0000-0002-3234-9368)
- Fabrice Roncoroni (ORCID: https://orcid.org/0000-0001-6402-3752)
- Declan P. McCarthy (ORCID: https://orcid.org/0000-0002-6381-3119)
- Maher B. Alghalayini (ORCID: https://orcid.org/0000-0002-4071-0272)
- Brett A. Helms (ORCID: https://orcid.org/0000-0003-3925-4174)
- Arian Gashi
- Oleg Kostko (ORCID: https://orcid.org/0000-0003-2068-4991)
- Edward S. Barnard (ORCID: https://orcid.org/0000-0003-4736-0743)
- Tim Kodalle (ORCID: https://orcid.org/0000-0002-8792-9669)
- Alexander Hexemer (ORCID: https://orcid.org/0000-0002-5269-0125)
- Morgan K. Wall
- Carolin M. Sutter‐Fella (ORCID: https://orcid.org/0000-0002-7769-0869)
- Yunfei Wang
- Chenhui Zhu
Institutions
- Lawrence Berkeley National Laboratory (US)
- Advanced Light Source (US)
- Molecular Foundry (US)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-18
- DOI
- https://doi.org/10.26434/chemrxiv.15012421/v1
- Primary Topic
- Machine Learning in Materials Science
- Type
- preprint
Funders
- U.S. Department of Energy
- Office of Science
- Basic Energy Sciences
- Lawrence Berkeley National Laboratory