Autobot platform uses machine learning to quickly find the best way to make advanced materials

A research team led by the Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab) has successfully built and demonstrated an automated experimental platform to optimize the production of advanced materials. A platform called Autobot uses machine learning algorithms to guide robotic devices to rapidly synthesize and characterize materials. The algorithm automatically improves the experiment based on analysis of characterization results.

Researchers tested the platform with an emerging class of material called Metal Halide Perovskites, which shows promise for applications such as optically emitting diodes (LEDs), lasers and photosectors. It took AutoBot weeks to investigate numerous combinations of manufacturing parameters to find the combinations that produce the highest quality materials.

Autobot, informed by machine learning algorithms with very fast learning rates, had to experimentally sample only 1% of the 5,000 combinations to find this “sweet spot.” This process takes up to a year with a traditional trial and error approach. In this approach, researchers manually tested a set of parameters that were guided by previous experience and intuition at once.

“Autobot represents a paradigm shift in material research and optimization,” said Carolin Sutter-Fella, a scientist at Berkeley Lab and one of the research's corresponding authors. “By integrating synthesis, characterization, robotics and machine learning capabilities on a single platform, Autobot dramatically accelerates the synthetic recipe screening process. Its rapid learning approach is a critical step towards establishing an autonomous optimization laboratory, allowing it to expand to a wide range of materials and devices.”

Molecular Foundry Scientist – User Facilities Bureau, Energy Science Division, Berkeley Lab – recognized the idea of ​​AutoBot, extended with a commercial robotics platform, implementing solutions for data processing, analysis, and machine learning infrastructure.

The interdisciplinary team included researchers from the University of Washington University, University of Nevada, University of California, University of California, University of California, Berkeley, Friedrich-Alexander-Universita Toerlangen – Nürnberg.

Scientists report their work in journals Advanced energy materials.

Iterative Learning Loop

Metal halide perovskites are highly sensitive to humidity, so strict air control is required to create high-quality thin films. As a result, it is difficult to implement cost-effective industrial-scale manufacturing. Using Autobot, the team identified synthetic conditions that allow for the production of high-quality thin film materials in high-quality humidity environments, addressing key barriers to large-scale production.

Autobot repeated a series of tasks, automatically adjusting tasks based on analysis of the results. This iterative learning loop proceeded as follows:

1. Autobot synthetic halide perovskite membranes changed four synthetic parameters from the chemical precursor solution. This is the timing to treat the solution with a crystallizer. Heating temperature; Heating time; Relative humidity of the film deposition chamber.
2. The platform characterized the samples using three methods. Measurement of ultraviolet and visible light passing through the sample (UV-vis spectroscopy). Light is applied to them and the emitted light is measured (photoluminescence spectroscopy). Emitted light is used to generate an image of the sample to assess the uniformity of the thin film (photoluminescence imaging).
3. The data workflow extracted information from characterization results, analyzed and combined the data into a single score representing film quality.
4. Based on these scores, the machine learning algorithm modeled the relationship between the synthesis parameters and film quality and determined the following experiments: These decisions were made with the aim of assessing the most beneficial parameter combinations to maximize the gain of the information at each iteration. This allows for efficient and accurate prediction of thin film material quality for all parameter combinations.

https://www.youtube.com/watch?v=-wy-7styw7u

Ultra-fast learning

Autobot has discovered that high quality films can be synthesized at relative humidity levels of 5% to 25% by carefully tuning the other three synthesis parameters.

“This humidity range does not require strict environmental management,” says Ansman Halder, a postdoctoral researcher at Berkeley Lab and co-first author of the research paper. “This discovery lays an important foundation for the development of commercial manufacturing facilities.”

Another insight was that humidity levels above 25% destabilizing the material during the deposition process, resulting in poor film quality. The team explained and validated this finding by performing manual photoluminescence spectroscopy during film integration.

Autobot's performance was impressive. By identifying the most useful experiments, the algorithm quickly learned how the synthetic parameters affect film quality.

“This powerful performance was demonstrated by a dramatic decline in the learning rate of the algorithm after Autobot sampled less than 1% of more than 5,000 parameter combinations,” said Maher Alghalayini, a postdoctoral scholar and co-author at Berkeley Lab. “In the new experiment, we decided to stop running the experiment because we had not changed the material quality predictions of the algorithm at this point.”

An innovative aspect of this research was the “multimodal data fusion.” This was done using a variety of data science and mathematical tools to integrate different data sets and images from three characterization techniques into a single metric for material quality. The idea was to quantify the results and make them usable by machine learning algorithms. For example, a collaborator at the University of Washington designed an approach to convert photoluminescent images into a single number based on how light intensity changes across images.