Designing sensors to detect trace amounts of chemicals — whether environmental toxins, disease biomarkers, or other compounds — has traditionally required years of meticulous trial and error. But researchers at the University of Chicago Pritzker School of Molecular Engineering (UChicago PME) and Argonne National Laboratory have developed a faster, smarter approach.
The researchers developed an AI model called a spiking graph neural network, which processes information in a way that loosely mimics how neurons in the brain send signals. Image Credit: Junhong Chen Lab
Using an artificial intelligence (AI) model inspired by the human brain, they’ve created a more efficient method for designing chemical sensors.
This is an exciting step toward being able to automate the discovery process.
Junhong Chen, Crown Family Professor, Molecular Engineering, Pritzker School of Molecular Engineering, University of Chicago
As a proof of concept, the team used their AI model to identify chemical probes for detecting PFAS, the so-called “forever chemicals”, in water. Early computer simulations suggest that the AI-selected probe could outperform existing PFAS sensors.
A Design Challenge
Many chemical sensors today use a tiny device known as a field-effect transistor (FET). These sensors feature a probe on the transistor’s surface that responds when it contacts a specific chemical, like a water pollutant, altering the transistor's electrical signal. That change lets scientists know the chemical is present.
While the concept sounds straightforward, building a functional FET sensor is anything but. The process typically requires selecting a probe material that binds only to the target chemical, and not to similar substances, while also functioning reliably under varying environmental conditions. With thousands of potential materials to choose from, evaluating each one in the lab is a time-consuming process.
With these sensors, you can tweak their design just a little bit to completely change the chemical probe and use it to detect different contaminants. But the complexity makes it challenging to pinpoint the optimal design for any given chemical.
Rui Ding, Eric and Wendy Schmidt AI in Science Postdoctoral Fellow and Study Co-First Author, University of Chicago
To accelerate the process, the UChicago PME team collaborated with Yuxin Chen from the Department of Computer Science and Claire Donnat from the Department of Statistics to develop an AI-driven solution. They created a spiking graph neural network, a type of model that processes information in a way loosely inspired by how neurons transmit signals in the brain.
These spiking graph neural networks have been successfully used to address problems in robotics and audio processing, but this is the first time they’ve been applied to chemistry or molecular engineering.
Rodrigo Ferreira, PME Graduate Student and Study Co-First Author, University of Chicago
The model was trained on data extracted from over a thousand published scientific papers, allowing it to “learn” the key characteristics of an effective chemical sensor. Armed with that knowledge, it could quickly and accurately, about 90 % of the time, predict which material combinations were most likely to succeed.
Sensing “Forever Chemicals”
To test the AI’s real-world performance, the researchers challenged it with a tough assignment: design a sensor for per- and poly-fluoroalkyl substances (PFAS). These synthetic compounds, used in everything from nonstick cookware to firefighting foam, are notoriously difficult to break down, earning them the nickname “forever chemicals.”
Detecting PFAS in water poses a serious challenge. The molecules are small, chemically slippery, and often resemble other substances, making them hard to isolate. To ensure a true test, the team deliberately excluded any PFAS-specific data during the AI’s training phase. The model was then asked to evaluate known probe materials and predict which might be most effective for detecting PFAS.
“We have some current probes that detect PFAS, but they do so with inadequate selectivity; they aren’t good enough yet. We thought that this model might be able to point us in new directions,” added Junhong Chen.
And it did. The AI identified graphene, a familiar material made of a single layer of carbon atoms, as a viable candidate. But it also flagged a less conventional option: ferrocenecarboxylic acid. Computer simulations showed that this pairing could potentially outperform existing sensors, particularly in how precisely it targets PFAS without interference from similar compounds.
The team has now begun experimental validation in the lab. If the simulations hold up, this approach could lead to faster, more accurate sensor development for a range of uses, from environmental monitoring to medical diagnostics.
Looking ahead, the researchers plan to expand the AI’s training data to further improve its accuracy across a wider range of sensor types. Their long-term vision? AI models that can autonomously design, plan, and even carry out experiments — dramatically reducing the time and labor involved in sensor development.
“This is an exciting initial trial to show that we can actually automate the sensor discovery process by leveraging large amounts of literature combined with our own expertise in this area. This could save a lot of human effort,” said Chen.
Funding for this research was provided by an Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, the University of Chicago Data Science Institute, the National Science Foundation (grant number 2037026), and Robust Intelligence (grant number 2313131).
Journal Reference:
Ferreira, P, R., et al. (2025) Expediting field-effect transistor chemical sensor design with neuromorphic spiking graph neural networks. Molecular Systems Design & Engineering. doi.org/10.1039/D4ME00203B