Science & Technology (Commonwealth Union) – The ability of artificial intelligence (AI) to analyze data and take action in a precise manner still has restrictions. However, as times goes by and AI improves the efficiency is likely to improve with human over sights.
Taking into account that machine learning cannot replace human expertise, but it can work tirelessly at a scale that humans cannot match, making it a valuable tool for discovering new chemicals and materials. Researchers already understand that machine-learning models can use enormous datasets to predict potential structures. The bigger question is whether AI can go beyond prediction and dramatically expand the process of experimentally testing those ideas.
Zhiling Zheng, an assistant professor of chemistry in Arts & Sciences at Washington University in St. Louis indicated that they already see that AI is powerful in terms of predicting new structures. He pointed out however that for most bench chemists and materials scientists, what matters more is actually producing those materials.
In a recent prize-winning essay published in the journal Science, Zheng outlined how AI could help address this next challenge. Zheng highlighted the fact that at the heart of this platform is the AI’s ability to read chemistry like a chemist.
Christopher Cooper, an assistant professor of energy, environmental and chemical engineering at the McKelvey School of Engineering, shares this perspective. In a recent paper published in Matter, he described methods for organizing and curating vast amounts of data related to polymer synthesis.
Data curation is the first crucial stage of the process. Researchers must gather the relevant information and transform it into a format that machine learning systems can readily interpret and process.
The models have already been trained on the established principles of chemistry, which allows them to generate predictions. However, what they lack is the ability to put those principles into practice by simulating how the predicted molecules could actually be produced. To bridge that gap, the machines need access to the “recipes” for chemical synthesis—the step-by-step procedures scattered throughout decades of scientific journals, textbooks and even footnotes. These instructions need to be collected, converted into usable data and supplied to the models.
Cooper pointed out that the goal is for a model to understand those instructions, combine them and determine which approach has the greatest likelihood of succeeding.
Zheng’s research background is rooted in metal-organic frameworks (MOFs), materials constructed from metal ions that serve as nodes and organic molecules that act as linkers. Their structures can resemble a cube, with metal components positioned at the corners and organic building blocks connecting them. This modular design offers enormous flexibility and possibilities for modification, much like a LEGO system for chemical synthesis. However, that flexibility also creates a major challenge: there are simply too many possible combinations to explore.
Zheng indicated that there are just so many different possibilities. The number of potential MOF designs could reach into the millions, making it practically impossible for researchers to experimentally test every option.
The concept of a self-driving or autonomous laboratory has existed in academic research for decades, Zheng explained. What has changed is the emergence of large language models (LLMs), which have made it possible to develop a new approach that was previously difficult to achieve. Researchers can train AI systems in a way similar to mentoring a graduate student—providing extensive instructions, allowing the system to practice, and enabling it to learn from errors along the way.
To demonstrate this approach, Zheng and his team trained LLMs using a literature-derived dataset containing around 4,000 MOF linker transformations. In simple terms, this represents roughly 4,000 different sets of building instructions. The researchers then trained an AI agent to narrow those possibilities by applying chemical constraints. Using computational simulations, the resulting design agents identified 10 new viable materials that showed better water-harvesting performance than leading aluminum-based adsorbents.
