Scientists are striving to explore new semiconductor materials that can enhance the efficiency of solar cells and other electronic devices. However, the pace of innovation has been constrained by the speed at which researchers manually measure key material properties.

Researchers at the Massachusetts Institute of Technology (MIT) have developed a fully automated robotic system that can accelerate this process.

Their system uses a robotic probe to measure an important electrical property called photoconductivity—the degree to which a material responds electrically to the presence of light.

The researchers infused domain knowledge from human experts in materials science into a machine learning model to guide the robot's decisions. This enables the robot to identify the optimal locations for probe contact with the material to obtain the most information about its photoconductivity, while a specialized path-planning program finds the fastest route to move between contact points.

In a 24-hour test, the fully automated robotic probe performed more than 125 unique measurements per hour, achieving greater precision and reliability than other AI-based methods.

By dramatically increasing the speed at which scientists can characterize important properties of new semiconductor materials, this approach could advance the development of solar panels that generate more electricity.

"I find this paper very exciting because it provides a pathway for autonomous, contact-based characterization methods. Not all important properties of materials can be measured in a contactless way. If you need to make contact with the sample, you want it to happen quickly and you want to maximize the amount of information you get," said Tonio Buonassisi, professor of mechanical engineering and senior author of the paper on autonomous systems.

His co-authors include first author and graduate student Alexander (Aleks) Siemenn; postdoctoral fellows Basita Das and Kangyu Ji; and graduate student Fang Sheng. The research will be published in Science Advances.

Making Contact

Researchers in Buonassisi's Accelerating Materials Development for Sustainability Lab are working to build a fully autonomous materials discovery lab. They have recently focused on discovering new perovskites, a type of semiconductor material used in photovoltaic technologies such as solar panels.

In previous work, they developed techniques for rapidly synthesizing and printing unique combinations of perovskite materials. They also designed imaging-based methods to determine some important material properties.

But photoconductivity can only be characterized by placing a probe on the material, shining light on it, and measuring the electrical response.

"To make our lab run as quickly and accurately as possible, we had to figure out a solution that yields the best measurements while minimizing the total time required for the entire process," Siemenn said.

Doing so required integrating machine learning, robotics, and materials science into an autonomous system.

First, the robotic system uses its onboard camera to capture an image of a glass slide printed with perovskite materials.

Then, it uses computer vision to segment the image into pieces, which are fed into a specially designed neural network model incorporating domain expertise from chemists and materials scientists.

"These robots can increase the reproducibility and precision of our operations, but importantly, humans still need to be in the loop. If we don't have a good way to apply the rich knowledge of these chemistry experts to our robots, we won't discover new materials," Siemenn added.

The model uses this domain knowledge to determine the best points for probe contact based on the sample's shape and material composition. These contact points are fed into a path planner that finds the most efficient route for the probe to reach all points.

The adaptability of this machine learning approach is particularly important because printed samples have unique shapes, from circular droplets to gummy bear-like structures.

"It's almost like measuring snowflakes—it's hard to find two that are exactly the same," Buonassisi said.

Once the path planner finds the shortest route, it sends signals to the robot's motors, which maneuver the probe to make measurements rapidly and consecutively at each contact point.

The speed of the method stems from the self-supervised nature of the neural network model, which determines optimal contact points directly on sample images without labeled training data.

The researchers also accelerated the system by augmenting the path-planning program. They found that adding a small amount of noise or randomness to the algorithm helps find the shortest path.

"As we move forward in this era of autonomous labs, you really need these three areas of expertise—hardware building, software, and understanding of materials science—coming together in the same team to innovate quickly. And that's one of the secrets," Buonassisi said.

Rich Data, Fast Results

After building the system from scratch, the researchers tested each component. Results showed that the neural network model could find more accurate contact points with shorter computation time compared to seven other AI-based methods. Additionally, the path-planning algorithm consistently found shorter routes than other approaches.

When they assembled all components for a 24-hour fully automated experiment, the robotic system performed over 3,000 unique photoconductivity measurements at a rate exceeding 125 per hour.

Moreover, the level of detail provided by this precise measurement method enabled the researchers to identify hotspots with higher photoconductivity as well as regions of material degradation.

Siemenn said: "Being able to collect such rich data at this speed without human guidance opens the door to discovering and developing new high-performance semiconductors, particularly for sustainable applications like solar panels."

The researchers hope to continue building this robotic system, working toward a fully autonomous materials discovery lab.