Yongtao Liu

Ancient civilizations are defined by the predominant materials used to make tools, i.e., Stone Age, Bronze Age, Iron Age, indicating the importance of materials for life and civilization. Until the late 20th century, humans mainly used natural materials. In the last century, materials science has rapidly emerged as one of the most important research fields, resulting in a diverse array of human-made materials. Today’s technological innovations in many cases involve or even depend on new materials. However, materials development is a trial-and-error process that is often laborious and time-consuming. A well-known example is the development of filaments for light bulbs. Prior to the first commercialized light bulb, more than six-thousand materials (e.g., carbon, platinum wire, etc) were tested as the filament, and the entire development and optimization process continued for more than a hundred years.

The rapid development of artificial intelligence and machine learning in recent years has led the materials science community to consider autonomous experiments through combining machine learning and instrument automation to accelerate the pace at which we understand and develop new materials. Automated and autonomous experiments can be carried out without human scientists’ intervention and consistently run with a high degree of accuracy. Analogous to classical experiments performed by scientists, autonomous experiments also require logical analysis of results and a repeatable procedure, such as operating scientific instruments and planning future experiments. Specialized devices, such as computer-controlled automatic systems, can be used to operate instruments. However, analyzing results and planning the next experiment typically rely on human intelligence. Therefore, a computer program that can simulate human intelligence—artificial intelligence—is a necessary part of autonomous experiments.

Figure 1. A scheme showing the working process of machine learning-driven autonomous scanning probe microscopy, where the microscope performs measurement and transfers results to the machine learning algorithm, then the machine learning algorithm performs real-time analysis and makes a decision for the next measurement accordingly. Based on this decision, the workflow will drive the microscope to perform the next measurement.

Recently, scientists who work on the Energy Frontier Research Center for 3D Ferroelectric Microelectronics (3DFeM) project at Oak Ridge National Laboratory developed a self-driving autonomous scanning probe microscope for materials investigation and discovery.1 Scanning probe microscopy is a powerful tool capable of ‘touching’ materials at the nanoscale level to ‘feel’ materials’ properties. The researchers developed a machine-learning algorithm that learns the relationship between image structure data and local spectroscopic data and then commands the microscope for the next measurement autonomously, allowing the microscope to quickly navigate over complex materials to identify structures and properties of interest, as shown in Figure 1. For instance, the autonomous scanning probe microscope can scan a ferroelectric material (for those who are interested in ferroelectric materials, please see this previous EFRC newsletter article2) for its domain structures, then navigate over the scanned area to explore the property related to each domain structure via spectroscopic measurement, while the active learning algorithm learns the structure–property relationship during experiments. This autonomous microscope dramatically increases the speed of discovering and understanding the fundamental principles that govern the behavior of materials.

The development of autonomous experiments in scanning probe microscopy is an exciting step in the materials science field due to the potential to accelerate the understanding of materials physics. Noteworthily, even if we implemented the above machine learning-driven autonomous experiment workflow in scanning probe microscopy, the machine learning algorithm and designed workflow can also be adapted for other experiments, such as, electron microscopy, chemical imaging, and material synthesis. The authors of this work1 are now working on improving the algorithm and making it available for a wide range of experiments in addition to scanning probe microscopy. Autonomous experiments can potentially revolutionize the field of materials science, carrying out experiments without the presence of human researchers to identify materials’ suitability for specific applications with high accuracy and efficiency.

More Information

[1] Liu, Yongtao, Kyle P. Kelley, Rama K. Vasudevan, Hiroshi Funakubo, Maxim A. Ziatdinov, and Sergei V. Kalinin. "Experimental discovery of structure–property relationships in ferroelectric materials via active learning." Nature Machine Intelligence 4, no. 4 (2022): 341-350.

[2] https://www.energyfrontier.us/content/unlocking-power-ferroelectrics-microelectronics



The development of the machine learning workflows was supported by the Oak Ridge National Laboratory’s Center for Nanophase Materials Sciences (CNMS), a US Department of Energy, Office of Science User Facility (M.A.Z., R.K.V.). The deployment of the machine learning workflows on the operational microscope was supported as part of the center for 3D Ferroelectric Microelectronics (3DFeM), an Energy Frontier Research Center funded by the US Department of Energy (DOE), Office of Science, Basic Energy Sciences under award no. DE-SC0021118 (Y.L., K.P.K., S.V.K.).

About the author(s):

Yongtao Liu is a postdoc in Center for Nanophase Materials Science at Oak Ridge National Laboratory and a member of the EFRC Center for 3D Ferroelectric Microelectronics. His research focuses on machine learning-driven autonomous scanning probe microscopy for materials discoveries. ORCID ID # 0000-0003-0152-1783.