Roll Out! AutoBot Is Transforming Research

Those of us that spend a large percentage of our time in the digital realm are spoiled by the rapid pace of technological progress. Things move really fast in the world of software, and with recent advances in artificial intelligence, we now even have software writing more software. But everything cannot be digitized, so this breakneck pace of innovation is all but impossible for more physical fields, such as scientific research.

This fact has not stopped a team led by researchers at the Lawrence Berkeley National Laboratory from trying to speed up scientific research, however. They have built a platform called AutoBot that leverages robotics and machine learning to automate the process of running experiments and evaluating the results. In particular, AutoBot can synthesize and evaluate materials to streamline the process of producing a novel material with desirable qualities. And through smart decision-making, AutoBot can squeeze a year’s worth of work into just a few weeks.

The team first demonstrated AutoBot’s potential by applying it to metal halide perovskites, a class of materials with promise for applications ranging from LEDs and lasers to photodetectors. These materials are notoriously difficult to work with, particularly because they are sensitive to humidity, which makes large-scale, cost-effective manufacturing difficult. AutoBot was tasked with finding conditions that would still yield high-quality thin films under less-than-ideal humidity.

Instead of testing all 5,000-plus possible combinations of fabrication parameters, AutoBot only needed to sample about 1% before it could make accurate predictions about the rest. Using machine learning to guide each iteration, the platform automatically adjusted synthesis parameters, characterized the results, and fed that data back into its decision-making loop. Within weeks, it had identified the sweet spot for producing reliable perovskite films — a process that would have taken human researchers up to a year using conventional methods.

One technique that made this possible was multimodal data fusion. The system combines different types of measurements, such as spectroscopy and photoluminescence imaging, into a single quality score for each sample. That unified score gives the machine learning algorithms a simple, usable metric for comparing results and deciding which experiment to run next. By quantifying complex optical and imaging data into a number, AutoBot essentially teaches itself to recognize good versus bad outcomes and improve with each iteration.

In addition to just automating work, AutoBot also uncovered some scientific insights along the way. For example, it found that high-quality perovskite films can still be produced in environments with relative humidity between 5% and 25%, so long as the other synthesis parameters are tuned carefully. This insight could eliminate the need for costly, ultra-dry manufacturing environments in the future.

As autonomous labs powered by platforms like AutoBot mature, they could revolutionize how quickly we discover and optimize materials, pushing the physical sciences a little closer to the rapid pace of progress we have come to expect in the digital realm.