AI helps scientists design better sensors for pollutants and beyond

Designing a sensor to detect miniscule levels of a chemical — whether it is an environmental toxin, a disease biomarker, or something else — has typically involved years of trial and error. Now, researchers at the University of Chicago Pritzker School of Molecular Engineering (UChicago PME) and Argonne National Laboratory have developed a powerful new method for designing such chemical sensors, using an artificial intelligence (AI) model mimicking the human brain.

“This is an exciting step toward being able to automate the discovery process,” said Junhong Chen, the Crown Family Professor of Molecular Engineering, Lead Water Strategist at Argonne, and senior author of the new paper, published in and featured on the cover of the journal Molecular Systems Design & Engineering (MSDE).

As a proof of principle, the team used the AI model to identify promising chemical probes for detecting PFAS, or “forever chemicals,” in water. Based on initial computer simulations, the AI-optimized chemical probe may work better than any existing PFAS sensor probes.

A design challenge

Many modern chemical sensors rely on a tiny device called a field-effect transistor, or FET. At their core, FET sensors contain a probe situated on the surface of a transistor – a tiny electronic switch. When a target chemical touches the probe — say, a pollutant in water — it changes the way electricity flows through the transistor. That change can be measured, letting scientists know the chemical is present.

The idea sounds simple, but building an effective FET sensor is challenging, often requiring many rounds of trial and error. Scientists must choose the right probe material, binding only to the chemical they’re looking for and not to similar molecules. They also must consider how that material behaves in real-world conditions, like in water with varying pH or temperature. There are thousands of possible probe materials, and testing them all in the lab would take forever.

“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,” explains Rui Ding, an Eric and Wendy Schmidt AI in Science Postdoctoral Fellow and co-first author of the paper. “But the complexity makes it challenging to pinpoint the optimal design for any given chemical.”

To speed things up, the UChicago PME team — in collaboration with Yuxin Chen of the Department of Computer Science and Claire Donnat of the Department of Statistics — turned to AI. They 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.

“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,” said Rodrigo Ferreira, a UChicago PME graduate student and co-first author of the new work.

The model was trained using data mined from more than a thousand past scientific papers, helping it “learn” what makes a good chemical sensor. With that knowledge, it could quickly predict which combinations of materials were most likely to work — and do it with about 90% accuracy.

Sensing “forever chemicals”

To see how well their AI model worked in practice, the researchers gave it a real-world challenge: designing a sensor for per- and polyfluoroalkyl substances (PFAS). PFAS are a group of man-made chemicals found in everything from nonstick cookware to firefighting foam, and they’ve earned the nickname “forever chemicals” because they don’t easily break down in the environment or the human body.

Detecting PFAS in water is notoriously difficult. The chemicals are small, slippery, and often hard to distinguish from other substances. Importantly, the team had not included PFAS-related data when training the AI, making it a true test of the model’s predictive power. They asked the AI to scan through a list of known probe materials and predict which ones might be best for detecting PFAS.

“We have some current probes that detect PFAS, but they do so with inadequate selectivity; they aren’t good enough yet,” said Chen. “We thought that this model might be able to point us in new directions.”

The AI highlighted one well-known material — graphene, a single layer of carbon atoms — but also flagged a less familiar candidate: ferrocenecarboxylic acid. The researchers ran detailed computer simulations and found that this new combination could potentially outperform existing sensors, especially in how well it zeroes in on PFAS without picking up unrelated chemicals.

They are now beginning real lab experiments to further test the results. If the results hold up, the method could open the door to designing better, faster sensors for all kinds of applications — from environmental safety to medical diagnostics.

In the future, the research team would like to continue expanding the data used to train the AI model, making it more accurate for a wider variety of sensor designs. Ultimately, they envision a future where AI models like this could autonomously design, plan, and even carry out experiments to discover new sensors with minimal human intervention.

“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,” said Chen. “This could save a lot of human effort.”

Citation: “Expediting field-effect transistor chemical sensor design with neuromorphic spiking graph neural networks,” Ferreira et al, MSDE, March 4, 2025. DOI: 10.1039/d4me00203b

Funding: This work was supported by an Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, the University of Chicago Data Science Institute, the National Science Foundation (2037026), and Robust Intelligence (2313131).