Artificial intelligence serves as a powerful ally for researchers but has a notable drawback: it often cannot clarify how it reaches its conclusions, which is referred to as the “AI black box.” Researchers from the University of Illinois Urbana-Champaign have successfully bridged this gap by integrating AI with automated chemical synthesis and experimental validation to uncover the chemical principles that AI used to enhance molecules for solar energy harvesting.
The research led to the creation of light-harvesting molecules that are four times more stable than their initial versions, alongside essential insights into the factors contributing to their stability — a challenging aspect in the development of materials.
This interdisciplinary effort was co-directed by several U. of I. professors, including Martin Burke in chemistry, Ying Diao in chemical and biomolecular engineering, Nicholas Jackson in chemistry, and Charles Schroeder in materials science and engineering, in partnership with Alán Aspuru-Guzik, a chemistry professor from the University of Toronto. Their findings were published in the journal Nature.
“Modern AI tools are incredibly effective. However, when you try to examine the underlying mechanisms, you typically find them opaque,” Jackson noted. “In the realm of chemistry, this can be quite frustrating. AI can assist in refining a molecule, but it often fails to explain why it’s the best option — what are the vital properties, structures, and functions? With our approach, we discovered what enhances the photostability of these molecules. We transformed the AI black box into a clear glass globe.”
The team aimed to tackle the problem of enhancing organic solar cells, which utilize flexible and thin materials as opposed to the conventional bulky, rigid silicon panels dominating present installations.
“The major obstacle in commercializing organic photovoltaics has been stability issues. High-performance materials deteriorate when exposed to light, which is counterproductive for solar cells,” Diao explained. “These materials can be produced and installed in ways not feasible with silicon, and they can also convert heat and infrared light to energy, but stability issues have persisted since the 1980s.”
The method developed in Illinois, termed “closed-loop transfer,” starts with a protocol for AI-guided optimization known as closed-loop experimentation. The researchers prompted the AI to enhance the photostability of light-harvesting molecules, according to Schroeder. The AI algorithm suggested various chemicals to synthesize and investigate over several iterations of closed-loop synthesis and experimental assessment. After each round, the new data was fed back into the model, leading to refined suggestions and progressively moving closer to the desired outcome.
Through this innovative approach, the researchers produced 30 new chemical candidates across five rounds of closed-loop experimentation, utilizing a modular, building block-like chemistry and automated synthesis pioneered by Burke’s group. This work was conducted at the Molecule Maker Lab within the Beckman Institute for Advanced Science and Technology at the University of Illinois.
“The modular chemistry technique works exceptionally well with the closed-loop experiment. The AI algorithm seeks new data that maximizes learning, while the automated molecule synthesis platform rapidly generates necessary compounds. Those compounds are then tested, with the data returning to the model, enhancing its intelligence — repeatedly,” Burke, who is also a professor at the Carle Illinois College of Medicine, remarked. “Previously, we primarily focused on structure. Our automated modular synthesis is now venturing into the realm of function exploration.”
In contrast to standard AI-driven processes that merely conclude with the final products identified by the AI, the closed-loop transfer approach aimed to reveal the underlying principles that contributed to the increased stability of the new molecules.
Throughout the closed-loop experiment, another set of algorithms was continuously assessing the created molecules, developing models that predict the chemical features linked to light stability, Jackson explained. Once the experiment was complete, these models generated new lab-testable hypotheses.
“We’re employing AI to produce hypotheses that we can verify to fuel new human-led discovery efforts,” Jackson stated. “Now that we have physical descriptors highlighting what gives molecules photostability, the screening process for identifying new chemical candidates becomes significantly easier, rather than aimlessly prowling through chemical space.”
To validate their hypothesis on photostability, the researchers explored three structurally distinct light-harvesting molecules with the identified chemical characteristic — a specific high-energy region — and confirmed that selecting the appropriate solvents could make these molecules up to four times more stable against light.
“This serves as a proof of principle for the potential of our approach. We are confident that we can tackle other material systems, with possibilities limited only by our imagination. Ultimately, we foresee a platform where researchers can input desired chemical functions and the AI will generate hypotheses to test,” Schroeder concluded. “This achievement was made possible through a collaborative, multidisciplinary team, leveraging the people, resources, and facilities at Illinois, as well as our collaborator in Toronto. Five groups united to yield new scientific insights that could not have emerged from any single subgroup working in isolation.”
This research received funding from the Molecule Maker Lab Institute, an AI Research Institutes program sponsored by the U.S. National Science Foundation under grant no. 2019897.
