Automating Innovation

Due to growing populations and the rise of technologies like artificial intelligence (AI), we can expect that our rate of energy consumption will continue to rise well into the future. In many regions, the existing energy infrastructure is already beginning to buckle under our demands. So in order to accelerate our rate of technological progress, alternative sources of energy will be necessary.

It has long been a dream to more fully utilize solar energy, in particular, due to its vast potential and sustainability. However, existing solar panels simply are not efficient enough to power our planet. Sure, they can play an important role in producing energy today, but to meet our growing needs and significantly reduce our reliance on fossil fuels and other less-than-ideal sources of energy, meaningful advances in solar technology will be required.

A group of researchers at MIT believes the best path forward is to provide materials science labs with tools to speed up their work. Oftentimes, these labs produce a large number of novel materials in an effort to find something with the right properties. But each of these materials must be manually evaluated, which is a slow and laborious process. The perfect material for ultra-efficient solar panels may already exist, but if it is lost in a sea of other options, it may go unnoticed.

To address this bottleneck, the researchers have developed a fully autonomous robotic system capable of dramatically accelerating the pace at which new semiconductor materials can be tested. This system centers around a robotic probe that measures photoconductance, which describes how well a material conducts electricity when exposed to light. Photoconductance is a key metric in evaluating materials for solar energy applications.

The system blends robotics with machine learning and expert knowledge from materials science to make intelligent, real-time decisions about where to measure each sample for maximum information gain. Using computer vision, the robot segments the sample, evaluates it using a custom neural network infused with expert insights, and selects the optimal contact points for probes. A specialized path planner then calculates the most efficient route between those points, ensuring high-speed operation without sacrificing precision.

During a 24-hour autonomous test, the robot performed over 3,000 unique photoconductance measurements (more than 125 per hour), demonstrating not only speed but also a level of accuracy and repeatability that surpasses other AI-driven methods.

The robot learns via a self-supervised approach, which means it does not require large amounts of labeled data to function. Instead, it learns directly from the structure and appearance of each new sample, adapting its strategy accordingly. This is particularly valuable in materials science, where samples are often irregularly shaped and vary significantly from one another.

In addition to speeding things up, the system also reveals detailed spatial maps of each sample’s electrical behavior. These maps allow researchers to find hotspots of performance and identify areas where the material may have defects or degradation. This information is very important for refining material formulations and manufacturing processes.

By removing the human bottleneck from materials testing, researchers can explore a vastly larger set of possible materials. This improves the odds of discovering the next generation of materials for use in solar panels and beyond.