AI/ML-Accelerated Analytics and Automation

Multi-fidelity Scale Bridging and Materials Design. CNM’s scientific staff, with diverse modeling and simulation expertise spanning high-fidelity electronic structure calculations to mesoscale models, aim to bridge the gap between the various simulation models’ accuracy and efficiency. This involves utilizing AI/ML algorithms, including symbolic regression, active and transfer learning, multi-objective evolutionary optimization, and reinforcement learning (RL), addressing spatiotemporal challenges in molecular and mesoscopic simulations. Material design can leverage such multi-fidelity models but often involve multi-objective, multi-dimensional search problems and learning domains that often have continuous search spaces. Conventional approaches rely on human intuition and metaheuristic searches, exhibiting issues like sluggish convergence and scalability concerns. Departing from traditional evolutionary, swarm, random sampling, and gradient-based methods, our proposed work involves developing machine learning approaches to efficiently navigate high-dimensional search landscapes, significantly improving search quality, convergence speed, and scalability in material discovery and design.

Theory-Guided Information Extraction from Experimental Characterization. Our objective is to efficiently extract information from multimodal experimental datasets, thus expediting scientific discovery, and utilize them for dynamical simulations or as training data to enhance models. In the spatiotemporal domain, a range of microscopy, spectroscopy, and scattering techniques enable the collection of spatiotemporally resolved information at varying resolutions. Recent efforts at the scientific user facilities (SUFs) have achieved ultrahigh spatial and temporal resolutions, providing unprecedented precision in visualizing processes at the electron and atomic levels. Various tools for imaging at scales from atomic to mesoscale resolution exist in the spatial domain, including transmission electron microscopy and coherent diffraction imaging. Innovative spatiotemporally resolved imaging tools, like the nanoprobe and UEM, have been developed, advancing material characterization. Despite these advancements, extracting information from spatiotemporal-resolved characterization remains a major challenge. Our approach involves developing AI/ML algorithms and workflows for swift analysis of extensive or high-dimensional multimodal data. This incorporates input from simulation models with embedded physics/chemistry models (physics-informed ML) to enable on-the-fly information extraction, facilitating real-time responses and accelerating scientific knowledge discovery.

Autonomous Synthesis of Nanoscale and Metastable Materials. At CNM, we aim to establish an AI-guided infrastructure empowering users to autonomously control metastability across diverse length scales—from single atoms to nanoscopic domains and mesoscale morphology. Traditional approaches, such as grid scans or experimenter intuition, fall short in identifying target-functional materials or revealing trends in such spaces. The field requires efficient prediction, exploration, and navigation of complex physics and material spaces. Key challenges include the need for modular robotics to facilitate on-site and on-the-fly sample transfer between synthesis and characterization. Addressing these challenges involves the development of autonomous experiment control, dynamically updating the data-retrieval strategy with each new measurement by incorporating rigorously encoded known physics to constrain the search space. Physics-informed autonomous experimentation would revolutionize materials physics, enabling the study of more ambitious material classes. Our proposed AI/ML tools for autonomous experimentation would provide real-time control of material synthesis, offering access to metastable, non-equilibrium materials achievable only through active control of the synthesis pathway. These autonomous platforms would facilitate studies on the synthesizability of complex pathway-dependent materials, contributing to fundamental research essential for developing next-generation digital manufacturing platforms relying on active computation and control of processing to produce desired materials.