Scientists are trying to discover new semiconductor materials that could increase the efficiency of solar cells and other electronics. But the rate of innovation is a bottleneck by the speed at which scientists can manually measure important material properties.
A fully autonomous robotic system developed by researchers MIT can accelerate.
Their system uses a robotic probe to measure the important electrical properties known as photochancja, i.e. electrically reacting material to the presence of light.
Scientists inject knowledge about learning materials from human experts to the machine learning model, which conducts a robot's decision. This allows the robot to identify the best places to contact the material with the probe in order to obtain the most information about its photouuction, while the specialized planning procedure finds the fastest way to go between contact points.
During the 24-hour test, a fully autonomous robotic probe performed over 125 unique measurements per hour, with greater precision and reliability than other artificial intelligence methods.
By a dramatic increase in the speed at which scientists can characterize the important properties of new semiconductor materials, this method can stimulate the development of solar panels that generate more electricity.
“I think that this article is extremely exciting because it is a path for autonomous methods of characterization based on contact. Not every important material property can be measured without contact. paper in the autonomous system.
His co -authors are the main author Alexander (Aleks) Siemnn, a graduate; Postdocs Basita Das and Kangyu Ji; And a graduate of Fang Sheng. Work appears today Scientific progress.
Making contact
Since 2018, scientists from the Buonassisi laboratory are working on a fully autonomous laboratory for discovering materials. Recently, they focused on discovering new Pervasts, which are a class of semiconductor materials used in solar solar, such as solar panels.
In previous works, they developed techniques of fast synthesis and printing of unique combinations of Pervian material. They also designed imaging methods to determine some important material properties.
But the photocopier is most accurately characterized by placing the probe on the material, light shine and measuring the electrical response.
“To enable our experimental laboratory to act as quickly and exactly as possible, we had to come up with a solution that would bring the best measurements, while minimizing the time needed to launch the entire procedure,” says Siemenn.
In this way, it required integration of machine learning, robotics and material science with one autonomous system.
At the beginning, the robotic system uses its on -board camera to make a image of a slide with the Perovian material printed on it.
Then he uses a computer vision to cut this image into segments that are transmitted to the neural network model, which has been specially designed to take into account the specialist knowledge of chemists and materials.
“These robots can improve the repetition and precision of our activities, but it is important to still have a man in a loop. If we do not have a good way to implement extensive knowledge from these chemical experts in our robots, we will not be able to discover new materials,” adds Siemenn.
The model uses this domain knowledge to determine the optimal points so that the probe can contact the shape of the sample and its composition of the material. These contact points are transferred to the planner of the path, which finds the most efficient way for the probe to reach all points.
The ability to adapt this approach to machine learning is particularly important because the printed samples have unique shapes, from a round drop to jelly -like structures.
“It's almost like measuring snowflakes – it is difficult to get two that are identical,” says Buonassisi.
When the planner of the path finds the shortest path, he sends signals to the robot engines that manipulate the probe and take measurements at each point of contact in fast succession.
The key to the speed of this approach is the self -sufficient nature of the neural network model. The model defines optimal contact points directly on the sample image – without the need for marked training data.
Researchers also accelerated the system, increasing the path to the path. They discovered that adding a small amount of noise or random to the algorithm helped him find the shortest path.
“As you progress in the era of autonomous laboratories, you really need all three of these specialist knowledge – building hardware, software and understanding of material sciences – meeting in the same team to be able to quickly introduce innovations. And this is part of the secret sauce here,” says Buonassisi.
Rich data, quick results
After building the system from scratch, scientists tested each component. Their results have shown that the neural network model found better contact points with less calculations than seven other AI methods. In addition, the path planning algorithm consistently found shorter path plans than other methods.
When they submitted all the elements to carry out a 24-hour fully autonomous experiment, the robotic system carried out over 3,000 unique photoconngling measurements at a rate exceeding 125 per hour.
In addition, the level of detail provided by this precise measuring approach enabled scientists to identify hotspots with a higher photoconmption, as well as areas of material degradation.
“The possibility of collecting such rich data that can be captured at such a fast pace, without the need for human tips, begins to open the door to be able to discover and develop new high -performance semiconductors, especially when using sustainable development, such as solar panels,” says Siemnn.
Scientists want to continue building this robotic system, trying to create a fully autonomous laboratory to discover materials.
These works are partly supported by First Solar, ENI through the myth of Energy Initiative, Mathworks, consortium Acceleration University of Toronto, US Energy Department and the National Science Foundation.
