While artificial intelligence-guided experimentation has emerged as a promising method of chemical discovery, using AI to uncover new chemical knowledge has remained largely explored. Published in the journal Nature, the new paper titled, “Closed-Loop Transfer Enables AI to Yield Chemical Knowledge” demonstrates that AI with a physics-based feature selection and supervised learning can provide fundamental chemical insights. In a collaborative effort, researchers from the University of Toronto and the University of Illinois at Urbana-Champaign leverage their new research methodology known as Closed-Loop Transfer, that combines experimentation and AI to not only optimize the discovery and development of new materials but also yield new understandings of chemical space.

“Everyone involved in this project was focused on this dream of AI designed molecules,” said Austin Cheng, a fourth-year Ph.D. student at the University of Toronto and key player in this paper. “At the time we started it, no one had done this before, and no one had put together such a team of diverse experts, so discovering the possibilities for closed-loop discovery was very exciting for everyone involved.”

What is Closed-Loop Transfer?

Closed-Loop Transfer (CLT) is a novel AI-guided methodology consisting of three phases. The first is closed-loop discovery, which integrates automated experiments with advanced modeling techniques. The process involves a continuous feedback loop where predictions made by AI inform experimental designs, and the results of these experiments, in turn, refine the AI models. This synergy allows researchers to explore vast chemical spaces efficiently and effectively. The second is to leverage the data gathered from this campaign along with physics-based descriptors to extract physical insights and make predictions across chemical space. The third and final phase is to experimentally validate these insights, yielding knowledge to inform future campaigns.

While CLT specifically focuses on chemical discovery and material optimization, the underlying principles of automation, feedback loops, AI integration, and data-driven insights make it similar to the Acceleration Consortium’s (AC’s) self-driving labs (SDLs), which focus on general materials discovery.

University of Toronto – Artificial Intelligence

At the forefront of the AI and modeling aspects of this project, the University of Toronto played a crucial role in applying the algorithms that drive the CLT approach. Led by AC Director Alán Aspuru-Guzik and Austin Cheng, the team leveraged Bayesian optimization algorithms previously developed by the Matter Lab which use physics-based features to make recommendations at each round of the loop. These models helped optimize towards highly photostable compounds, which are essential for various applications, including solar energy and electronic devices.

By employing Bayesian optimization, the Toronto team was able to balance exploration and exploitation within the chemical space, guiding the closed-loop experiments toward promising candidates. This innovative use of AI not only enhanced the efficiency of the discovery process by not requiring all of the potential chemical space to be explored, but also provided deeper insights into the fundamental properties of molecules.

“Our team was proud to contribute the AI-driven methodologies to the Closed-Loop Transfer approach,” said Dr Aspuru-Guzik. “Our focus on developing sophisticated Bayesian optimization algorithms has not only enhanced our ability to predict molecular properties but has also provided fundamental insights into chemical stability. This collaboration with Marty’s team exemplifies how integrating advanced AI with experimental science can accelerate discovery and innovation in materials chemistry.”

University of Illinois – Experimentation

Complementing the AI efforts, the University of Illinois took charge of the experimental and computational work. Researchers, such as AC member Martin D. Burke and Charles Schroeder conducted automated modular synthesis of small molecules, while Ying Diao characterized their properties through rigorous testing, all while informed by computational physical modelling by Nicholas Jackson. This hands-on experimentation was vital for validating the predictions made by the AI models and for generating empirical data that could inform future research.

“This project powerfully harnessed the highly synergistic strengths of the Molecular Maker Lab Institute at the U of I and the Acceleration Consortium at U of T, paving the way for advancements in solar energy and beyond,” said Dr Burke. “We leveraged frontier automated modular synthesis, leading expertise in organophotovoltaics and theoretical chemistry, and the powerful Gryffin algorithm for highly efficient Bayesian optimization. The combined team harmonized our strengths to enable AI to uncover new chemical knowledge.”

A Collaborative Success

“The way the two teams worked together fostered a sense of urgency and committedness to the project,” said Cheng. “We built the right culture where experts in each area guided the experts in every other area, from AI to synthesis to devices to theory, which made us even more driven and motivated to succeed for each other.”

The collaboration between the University of Toronto and the University of Illinois exemplifies the power of interdisciplinary research. By combining cutting-edge AI techniques with robust experimental methodologies, the team has not only optimized the search for new materials but has also uncovered valuable insights into molecular photostability.

This research highlights the potential of AI-guided approaches in scientific discovery, paving the way for future innovations in chemistry and beyond. As we continue to explore the capabilities of AI in various fields, the findings from this collaborative effort serve as a promising example of how technology can enhance our understanding of the natural world.

Read the full article in Nature: https://www.nature.com/articles/s41586-024-07892-1.

‍