Machine Learning Group
Human-AI-Facility Integration for the Multi-Modal Studies on High-Entropy Nanoparticles
This project is developing the extensible infrastructure necessary for Human-AI-interfaced autonomous experiments that integrate synthesis and characterization at a beamline. It is building on recent strides using Bluesky to orchestrate human-in-the-loop [semi-]autonomous multimodal experiments at National Synchrotron Light Source II beamlines, contemporary flow reactor design for high-throughput synthesis at the Center for Functional Nanomaterials (CFN), and artificial intelligence (AI)/algorithmic advancements led at CSI.
We are building the necessary software infrastructure to integrate an existing CFN-built flow reactor using Bluesky, as well as the user interfaces needed to drive a multimodal experiment through a human-AI collaboration. This will culminate in a study that combines diffraction and spectroscopy of functional high-entropy nanoparticles and optimizes the scientific understanding of their design space. The Human-AI-Facility, or HAI-FI, approach will ensure diverse datasets are interpretable to researchers in real time, enabling engagement with AI-driven experiments.
The team includes experts with expertise in materials synthesis, self-driving experiments, software engineering, and deploying AI at beamlines. The cross-directorate collaboration that takes advantage of this combined expertise will enable future user programs.
The three core tasks that make up this work include:
- Task 1, which involves using existing flow reactors to create an extensible abstraction through Bluesky’s Ophyd application programming interface (API).
- Task 2 consists of using this platform to design a framework for running real experiments using the reactors, driven and interpreted by a human-AI collaboration.
- Task 3 will focus on studying high-entropy materials in a live flow reactor, combining synthesis with existing multi-modal measurement capabilities and data management tools.
Recent work has showcased the benefits of using optimal experimental design (OED) for the orchestration of experiments at large-scale user facilities. These accomplishments suggest that AI-assisted experimentation capabilities have the potential to greatly accelerate exploration and targeted discovery of new complex compounds. In addition, seamlessly introducing human intuition and decision making in tandem with an autonomous agent can be used to great effect, giving experimenters the ability to leverage prior knowledge where possible and convenient. This HAI-FI style of techniques already has yielded dividends.
Publications
M. R. Carbone, H. J. Kim, C. Fernando, S. Yoo, D. Olds, H. Joress, B. DeCost, B. Ravel, Y. Zhang, P. M. Maffettone. Emulating Expert Insight: A Robust Strategy for Optimal Experimental Design. arXiv.2307.13871 (2023).
P. M. Maffettone, D. B. Allan, S. I. Campbell, M. R. Carbone, T. A. Caswell, B. L. DeCost, D. Gavrilov, M. D. Hanwell, H. Joress, J. Lynch, B. Ravel, S. B. Wilkins, J. Wlodek & D. Olds. Self-driving Multimodal Studies at User Facilities. arXiv:2301.09177 (2023). [Presented at the 36th Conference on Neural Information Processing Systems (NeurIPS 2022)].
