This machine learning model takes MOF simulations to the next level

A new study presented at the ACS Spring 2023 meeting shows how the powerful computational techniques behind artificial intelligence tools such as ChatGPT can help chemists train machines to better understand metal-organic frameworks (MOFs). The researchers shared their findings on Thursday, March 30, in a talk in the Division of Computers in Chemistry.

Chemists have trained neural networks to build machine learning models of MOFs that are capable of predicting specific properties of the material, such as electrical conductivity or its ability to capture volatile gases. But these neural networks generally can’t share between each other what they’ve learned from their task-specific training—partly because the properties of nanoporous materials are derived from both local features, such as the atomic composition of metal nodes and linkers, and global features, such as pore size.

“It would be nice to have a new, universal model that is trained not just for a specific task but [on the] general scope of what the MOF is like,” Jihan Kim, a computational chemist at the Korea Advanced Institute of Science and Technology, told C&EN. Kim and his colleagues were inspired by recent AI advances that make machine learning models better at interpreting multiple kinds of data using a neural network architecture called a transformer. Such capabilities are demonstrated by natural language processing models such as the tech company OpenAI’s ChatGPT, which generates text responses to a wide array of user prompts in a conversational style. So Kim wanted to see if transformers could give machines a more holistic handle on MOF chemistry.

Kim and his team pre-trained their model, called MOFTransformer, using a library of 1 million simulated MOFs that includes information about local and global features that human chemists consider when interpreting the materials’ physical properties. Using a second, smaller training set, the group then fine-tuned MOFTransformer to predict a variety of physical properties derived from different types of features. When they tested their model on IRMOF-1, a well-studied MOF structure, the researchers found that MOFTransformer accurately predicted the material’s hydrogen uptake, hydrogen diffusivity, and band gap properties. Further, the transformer showed the team exactly which features the model paid attention to when performing its calculations (Nat. Mach. Intel. 2023, DOI: 10.1038/s42256-023-00628-2). This could allow users to check the platform’s predictions against their own chemical intuition. Kim and his team hope this work will allow researchers to use MOFTransformer as the base model for their own investigations rather than having to start from scratch.

“It’s a nice application of really advanced machine learning models that [were] developed for other areas but applied to nanoporous materials,” said Tom Woo, a computational chemist specializing in MOFs at the University of Ottawa who was not involved in the study. He thinks the platform will be especially helpful for researchers trying to develop machine learning models with limited training data, he says. “It’s the next level of machine learning for nanoporous materials,” Woo said.