It’s not easy inventing green.
Sure, it’s easy to make a material green, orange, red or any other color, but to “inverse design” it, to work backwards from a specific color to design and build a material from the molecule up that hits the exact hue, tint, tone and shade, is one of the most persistent challenges in material synthesis.
“If you ask me for a formula that can give you a red color, that's no problem. I can write down a list of formulas for you,” said University of Chicago Pritzker School of Molecular Engineering (UChicago PME) Asst. Prof. Jie Xu, who has a joint appointment at Argonne National Laboratory. “But if you ask me, ‘Okay, I want a color that has this RGB value,’ that's a very precise color description, so that’s challenging.”
It takes decades to build up the know-how to inverse design for color or any other specific property, Xu said.
“It’s a grand challenge even for experienced material scientists,” she said. “If you have many years of experience in a field, you can probably do inverse design based on your knowledge.”
In new research from UChicago PME, Argonne and Purdue University, the AI-driven robotic lab Polybot hit the bullseye – designing and building new polymers that hit precise, targeted shades of orange and green – in just 72 hours.
The results were published in the Journal of the American Chemical Society.
While the results are remarkable, color is only a proof of concept. The researchers chose one of the most difficult color targets in polymer design to demonstrate their new machine learning framework, which they plan to extend to broader material goals beyond color.
“The way chemistry has been done has remained the same for centuries – you go to the lab, you set up a reaction in a flask, you heat the flask, then you work it up, you do the purification, etc.,” said co-author Jianguo Mei, the Richard and Judith Wien Professor of Chemistry at Purdue University. “Chemistry needs to embrace AI and automation. Chemistry and chemists need to actually change how things work.”
Hitting the target
The researchers chose to demonstrate their technique using color because of its difficulty, but also for its usefulness in electrochromic applications.
“Electrochromic polymers have been used widely for things like smart windows and tinted glasses, and have exciting future applications for virtual reality headsets and sustainability,” Xu said.
Orange and green were chosen as two of the less-used and therefore less-explored colors with fewer data points to feed to AI.
Rather than feeding it processed datasets, they trained the language model on full, published papers, teaching the AI to filter usable data out of the writing, illustrations, charts and graphic design flourishes of a magazine-style academic journal.
“The main objective is to demonstrate a general approach that combines literature mining, a self-driving lab and physics-informed predictive models to create efficient, accurate inverse design,” said co-author Henry Chan, a staff scientist at Argonne.
After it was fed just 19 papers, Polybot, housed in the Center for Nanoscale Materials, a DOE Office of Science user facility at Argonne, was running and rerunning experiments, modifying its calculations each time based on the previous results.
Within 72 hours, the team was holding plastic samples colored the exact orange and green they asked for.
Hitting all the targets
The team says their approach is applicable for hitting other targets in material design, for example surface morphology, stability, electrical and electrochemical performance. The next grand challenge is to hit many targets all at once.
“To optimize real-world materials, we have to look at multiple properties at the same time,” Chan said.
Taking the example of smart windows, which cut buildings’ energy costs by dimming, lightening or changing opacity on command. They need to be the right color, but to become useful real-world products, they also have to switch between light and dark quickly, be durable, stable, nontoxic and as resistant to heat and cold as a traditional storm window.
Chan envisions a system where a researcher specifies an overall goal for the self-driving lab to hit, then prorates material properties by importance.
“If we think that color is more important than the switching speed, then we may tell the self-driving lab to put 75% focus on color and 25% focus on the switching speed,” Chan said. “You define the priorities of the different objectives, and then let the system optimize the whole design space.”
Co-author Ian Foster, director of Argonne’s Data Science and Learning division and University of Chicago Arthur Holly Compton Distinguished Service Professor of Computer Science, said AI and humans working together will help advance science as a whole.
“In a sense, scientists, over decades or longer, have solved most of the easy problems. Only hard problems remain,” Foster said. “Humans, thinking very hard, will continue to make progress, but it's becoming harder and harder to make really transformational discoveries. AI methods turbo-charge the scientists as they seek to solve these complex problems.”
Citation: “Autonomous Synthesis and Inverse Design of Electrochromic Polymers with High Efficiency and Accuracy,” Wu et al, Journal of the American Chemical Society, November 21, 2025. DOI: 10.1021/jacs.5c12241