“`html
Researchers are endeavoring to uncover novel semiconductor substances that could enhance the efficacy of solar cells and various electronics. However, the velocity of advancement is hindered by the speed at which scientists can manually assess vital material characteristics.
An entirely automated robotic system created by MIT researchers may expedite this process.
This system employs a robotic probe to evaluate a crucial electrical characteristic referred to as photoconductance, which indicates how electrically reactive a material is when exposed to light.
The scientists incorporate materials-science expertise from human professionals into the machine-learning model that directs the robot’s decision-making. This allows the robot to pinpoint optimal locations for probing a material to extract maximum information regarding its photoconductance, while a specific planning process determines the quickest route between contact points.
During a 24-hour trial, the fully automated robotic probe recorded over 125 distinct measurements each hour, with greater accuracy and dependability compared to other AI-based techniques.
By significantly enhancing the pace at which researchers can characterize vital properties of new semiconductor materials, this approach could accelerate the creation of solar panels that generate more energy.
“I find this study to be immensely thrilling because it offers a pathway for autonomous, contact-based characterization techniques. Not every critical property of a material can be gauged without contact. When contact with your sample is required, you want the process to be swift and to maximize the information acquired,” states Tonio Buonassisi, a professor of mechanical engineering and lead author of a publication on the autonomous system.
His co-authors include lead writer Alexander (Aleks) Siemenn, a graduate student; postdoctoral researchers Basita Das and Kangyu Ji; and graduate student Fang Sheng. The findings are published today in Science Advances.
Establishing contact
Since 2018, the team in Buonassisi’s lab has been advancing towards a fully automated materials discovery laboratory. Their recent focus has been on discovering new perovskites, a category of semiconductor materials employed in photovoltaics such as solar panels.
In previous efforts, they developed techniques to rapidly synthesize and print unique combinations of perovskite material. They also devised imaging-based methods to ascertain some key material properties.
Nonetheless, photoconductance is most accurately evaluated by placing a probe on the material, illuminating it, and measuring the electrical response.
“To enable our experimental lab to function as quickly and accurately as possible, we needed to devise a solution that would yield the best measurements while reducing the total runtime of the procedure,” mentions Siemenn.
This required the synthesis of machine learning, robotics, and material science within a single autonomous system.
Initially, the robotic system utilizes its onboard camera to capture an image of a slide containing the printed perovskite material.
Then, it leverages computer vision to segment that image, which is input into a neural network model designed to incorporate domain expertise from chemists and materials scientists.
“These robots can enhance the repeatability and accuracy of our operations, but it remains vital to have a human in the process. Without effectively integrating the extensive knowledge of these chemical experts into our robots, we will fall short in uncovering new materials,” Siemenn further explains.
The model applies this domain knowledge to identify optimal points for the probe to contact based on the sample’s shape and material composition. These contact points are relayed to a path planner that determines the most efficient route for the probe to reach all locations.
The flexibility of this machine-learning method is particularly crucial since the printed samples possess unique shapes, from circular droplets to jellybean-like formations.
“It’s akin to measuring snowflakes — capturing two identical ones is nearly impossible,” Buonassisi states.
Once the path planner computes the shortest route, it transmits signals to the robot’s motors, which maneuver the probe and take measurements at each contact point in rapid succession.
Central to the speed of this method is the self-supervised aspect of the neural network model. It determines optimal contact points directly from a sample image — eliminating the need for labeled training data.
The researchers also expedited the system by refining the path planning process. They discovered that introducing a minor degree of noise, or randomness, into the algorithm aided in finding the most direct route.
“As we advance into this era of autonomous laboratories, it is essential to have expertise in hardware development, software, and a deep understanding of materials science converge within the same team to foster quick innovation. This integration is part of the key here,” Buonassisi remarks.
Abundant data, swift outcomes
After constructing the system from scratch, the researchers evaluated each component. Their findings demonstrated that the neural network model identified better contact points with reduced computation time than seven other AI-focused techniques. Additionally, the path planning algorithm consistently produced shorter routes than alternative methods.
When they combined all elements for a 24-hour fully autonomous experiment, the robotic system executed over 3,000 unique photoconductance measurements at a pace surpassing 125 per hour.
Moreover, the level of detail provided by this precise measurement technique enabled the scientists to pinpoint hotspots exhibiting higher photoconductance as well as regions of material deterioration.
“The ability to collect such rich data at such rapid rates, without requiring human intervention, begins to open avenues for discovering and developing new high-performance semiconductors, particularly for sustainable applications like solar panels,” states Siemenn.
The researchers aim to continue enhancing this robotic system as they work toward establishing a fully autonomous lab for materials discovery.
This research is partially funded by First Solar, Eni through the MIT Energy Initiative, MathWorks, the University of Toronto’s Acceleration Consortium, the U.S. Department of Energy, and the U.S. National Science Foundation.
“`