Researchers are working to identify new semiconductor materials that could enhance the performance of solar cells and electronic devices. However, progress is slowed by the time-consuming process of manually measuring key material properties.
Accelerating Discovery with Autonomous Photoconductance Testing
To accelerate this process, MIT scientists have created a fully autonomous robotic system. This system employs a robotic probe to assess photoconductance—a crucial electrical property that indicates how a material responds to light.
By integrating expert knowledge from materials science into the machine-learning model that directs the robot’s actions, the system can determine the most informative spots to probe. A custom planning algorithm then calculates the most efficient route between these points.
Over a 24-hour period, the fully autonomous robotic probe recorded over 125 distinct measurements each hour, delivering greater accuracy and consistency than other AI-driven techniques.
By significantly accelerating how quickly researchers can analyze critical characteristics of new semiconductor materials, this approach could help advance the development of more efficient solar panels.
“This paper is particularly exciting because it outlines a clear route for autonomous, contact-based measurement techniques,” says mechanical engineering professor Tonio Buonassisi, senior author of the study. Measuring all essential material properties often requires making physical contact. When contact is necessary, speed and data richness become crucial.
The research team includes lead author and graduate student Alexander (Aleks) Siemenn, postdoctoral researchers Basita Das and Kangyu Ji, and graduate student Fang Sheng. The researchers publish their findings today in Science Advances.
Establishing Contact
Since 2018, researchers in Buonassisi’s lab have been working toward creating a fully autonomous lab for materials discovery. More recently, their efforts have centered on identifying new perovskites—a type of semiconductor commonly used in solar panels.
In earlier projects, they developed methods for quickly synthesizing and printing unique perovskite combinations, along with imaging techniques to assess key material properties.
However, the most precise way to measure photoconductance involves placing a probe directly on the material, illuminating it, and recording its electrical reaction.
“To make our lab operate as efficiently and accurately as possible, we needed a solution that could deliver high-quality measurements while keeping the overall testing time to a minimum,” explains Siemenn.
This required combining machine learning, robotics, and materials science into a single autonomous system. The process starts with the robot using its built-in camera to capture an image of a slide containing the printed perovskite material.
Next, computer vision breaks the image into sections, which are analyzed by a neural network trained to integrate expert knowledge from chemists and materials scientists.
Blending Human Expertise with Robotic Precision for Smarter Measurements
“Robots can enhance the consistency and accuracy of our experiments, but human oversight remains essential. Without effectively embedding the deep insights of chemical experts into these systems, we won’t succeed in discovering new materials,” Siemenn notes.
The model applies this expert input to identify the best locations for the probe to make contact, based on the sample’s shape and composition. These contact points are then processed by a path-planning algorithm that calculates the most efficient route for the probe to follow.
This machine-learning method is especially valuable because the printed samples vary in shape—ranging from round droplets to irregular, jellybean-like forms.
“It’s a bit like trying to measure snowflakes finding two exactly alike is nearly impossible,” says Buonassisi.
Once the optimal path is identified, the system sends commands to the robot’s motors, which move the probe and rapidly collect measurements at each contact point.
The system’s speed largely comes from the neural network’s self-supervised learning. It selects the best contact points directly from the sample image, without requiring pre-labeled data.
The researchers also sped things up by improving the path planning algorithm. Interestingly, introducing a small degree of randomness allowed it to discover shorter, more efficient paths.
“As autonomous labs advance, it becomes essential to combine hardware engineering, software development, and materials science expertise within one team. That integration is a key ingredient to fast innovation,” Buonassisi explains.
Comprehensive Data, Swift Outcomes
After constructing the system from scratch, the researchers evaluated each of its components. They found that their neural network identified optimal contact points more efficiently and with less computational effort than seven other AI techniques. Additionally, the path planning algorithm consistently produced shorter routes compared to alternative methods.
When integrated into a fully autonomous 24-hour experiment, the robotic system performed over 3,000 distinct photoconductance measurements—averaging more than 125 per hour.
This high-precision approach also allowed the team to detect zones with elevated photoconductance and regions where the material had begun to degrade.
“Capturing such detailed data at high speeds and without human intervention paves the way for discovering and designing new, high-efficiency semiconductors—particularly for sustainable technologies like solar panels,” says Siemenn.
The team plans to further advance this robotic system as part of their goal to develop a fully autonomous materials discovery laboratory.
Read the original article on: MIT Massachusetts Institute Of Technolohy
Read more: Humanoid Robot Insists it has no Intention of ‘Replacing Human Artists’