A new pathway to custom functional polymers

At a potluck, you ate the best chocolate chip cookie—golden-brown, thick and chewy. Unfortunately, you don’t know who made the cookie to get the recipe from, so you decide to recreate it. Using forward design principles, you might randomly choose a recipe from dozens of options, bake and observe the resulting cookies. If they are too thin, you might start over with a new recipe, add more flour or chill the dough longer and make a new batch. An alternative method is to start from the cookie characteristics you want and ask: What recipe and baking settings will produce that type of cookie? This method is called inverse design.

That “work backward from the goal” mindset is increasingly important in modern materials science. Researchers often know what they want a material to do: conduct electricity better, tolerate heat, respond to a stimulus or display a very specific color. However, translating those requirements into a chemical formula and a reliable manufacturing method can take months or years of trial and error.

The challenge is especially steep for polymers, the long-chain molecules used in everything from packaging to medical devices to electronics. Polymers are built from smaller molecular “building blocks,” and small changes in those building blocks or how they are combined can lead to large and sometimes unpredictable changes in performance.

Autonomous inverse design for polymers

Researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory, the University of Chicago and Purdue University have now demonstrated a faster path: an autonomous inverse-design workflow that helps scientists go from a target property to a polymer recipe with far fewer experiments. Their paper is published in the Journal of the American Chemical Society.

“This work really marks a milestone in autonomous, on-demand production of functional materials,” said Jie Xu, a scientist from Argonne and an assistant professor at the University of Chicago Pritzker School of Molecular Engineering. “Instead of guessing and iterating for months, we can start from the property we want and let the system guide us to a polymer recipe.”

The approach connects three pieces that are often separate in traditional research. First, it gathers prior knowledge by automatically extracting data from published scientific papers, including information that appears in both text and images. To do this, the team used artificial intelligence (AI) “reading” tools—including large language models, like those used in today’s chatbots—to scan papers and pull out the details researchers normally collect by hand. These AI tools can identify and organize information buried in paragraphs, tables and even images.

“Modern AI can read the scientific record at a scale no person can, pulling out the key ingredients and results and turning them into usable data,” said Ian Foster, director of Argonne’s Data Science and Learning division, Argonne Distinguished Fellow and University of Chicago computer science professor. “That gives the lab a running start. And when the models learn from each new experiment, the system gets smarter as it goes.”

AI reading tools meet robotics

Second, the approach uses machine learning to predict which combinations of building blocks are most likely to produce the desired result. Third, predictions are sent directly into an automated laboratory workflow that can synthesize the polymers, purify them, prepare samples, measure their properties and feed the results back to improve the next round of predictions.

“Think of it like a GPS for chemistry,” said Argonne scientist Henry Chan. “Instead of wading through thousands of possible polymer recipes, the AI uses what’s already known to suggest the next best turn, then the robots quickly test it and report back.”

This workflow used Polybot, a self-driving laboratory platform housed in the Center for Nanoscale Materials, a DOE Office of Science user facility at Argonne. Designed to coordinate robots and instruments through an AI-driven system, Polybot allows experiments to run continuously with minimal human intervention. Polybot was not just used to automate a known procedure but to carry out an inverse-design loop where each experiment is chosen to move closer to the desired result.

Targeting precise electrochromic colors

To show how the workflow performs on a problem where precision matters, the team focused on electrochromic polymers, materials that can change color or transparency when a small voltage is applied. These polymers are candidates for technologies like smart windows that reduce building energy use and tinted layers for augmented- and virtual-reality headsets.

“Electrochromic polymers are especially compelling because they can be printed and patterned into device-ready formats,” said Materials Scientist Yuepeng Zhang, who leads the printable electronics lab at Argonne’s Materials Engineering Research Facility and worked with the team to print these functional polymers into prototype display devices.

“While scientists have created many electrochromic polymers over the years, hitting an exact shade of color is much harder than simply producing just ‘red’ or ‘green.’ Color can be quantitatively defined using standardized color coordinates, such as RGB values, but obtaining a close color match generally requires extensive optimization of the polymer recipe,” said Jianguo Mei, Richard and Judith Wien Professor of Chemistry at Purdue University, and a co-author of the study. RGB refers to the three-color video display model using red, green and blue light.

Building and exploring a color map

The researchers first built a dedicated electrochromic polymer database by mining published literature for polymer structures, synthesis details and data on the colors these materials produce. AI helped speed up the “knowledge-gathering” step by turning scattered information from many papers into a consistent, usable dataset, creating a searchable recipe-and-results library.

From the collected data, they calculated standardized color values and created a “map” of the colors that had already been reported. They then asked Polybot to aim for two specific color targets located in regions of the map where few or no polymer “recipes” had been documented: challenging shades of green and orange.

Within 72 hours, Polybot autonomously suggested polymer recipes, made the materials and checked the results against the targets. Even though there were more than a thousand possible compositions to choose from, the system narrowed in on close matches in only a few dozen experiments. It did this by making small, step-by-step adjustments to the ratios of three different building blocks. Each new result was added back into the database, helping the model make better predictions in the next round.

Expanding to broader material challenges

The significance extends beyond color-changing materials. This approach, combining automated knowledge gathering, predictive machine learning and iterative cycles of robotic experimentation, can be adapted to other material challenges where the search space is enormous, and the cost of trial-and-error is high.

As Xu and Mei put it, “Inverse design of electrochromic properties represents the first demonstrated capability enabled by our AI-robotic framework. The same approach can be extended to other properties, such as mechanical, optical and electrical characteristics, as well as to other classes of materials.”

By turning materials design into a faster, more systematic process, this work points toward a future where scientists can request a property and rapidly receive a tailored recipe, ready for further development and real-world applications.