A Johns Hopkins University engineer has developed a specialized AI tool that could do for materials scientists what ChatGPT has done for coders and writers. The new system, called ChatGPT Materials Explorer, or CME, could speed the discovery of everything from advanced batteries to tougher alloys, according to findings appearing in _Integrating Materials and Manufacturing Innovation_.
Image caption: Kamal Choudhary
"ChatGPT Materials Explorer is like having a specialized research assistant who is trained specifically to dig through huge databases, predict how a material or materials will behave without physical testing, sort through scientific papers to find studies relevant to your projects, and even analyze work and assist with scientific writing," says CME inventor Kamal Choudhary, a professor of materials science and engineering at the university's Whiting School of Engineering.
The tool's key innovation is its access to real scientific data and physics-based models, enabling it to give accurate answers to questions posed by materials scientists. Choudhary's experiences with ChatGPT inspired its creation.
"I work on a lot of superconductors, which are materials that conduct electricity without any resistance," says Choudhary, who also holds a joint appointment in the Department of Electrical and Computer Engineering. "I would ask ChatGPT, 'Can you design a superconductor with a particular composition and show me the crystal structure?' It gave me a very generic response, which turned out to be the wrong answer."
What Choudhary was experiencing are called hallucinations—when ChatGPT presents false information as factual, a not uncommon occurrence. Some experts estimate that ChatGPT has a hallucination rate of between 10% and 39%.
"Hallucinations happen because ChatGPT isn't trained to understand facts," Choudhary says. "If it can't find the exact answer based on the data it's pulling from, it will say something that sounds plausible. Data sources like Wikipedia or The New York Times don't often include current facts and research about materials science and can lead to incorrect answers. CME pulls its information from materials science databases, so its answers can be trusted by materials scientists."
Choudhary developed his specialized language model with the ChatGPT builder feature, which enables users to create custom GPTs tailored to their needs. He started by telling the AI broadly what he wanted it to do and setting parameters for its functions. Then he configured it, connecting the AI to the databases and instructing it on what kinds of answers it can give.
"These databases are how ChatGPT gets its information, so plugging in databases that are relevant to the field is crucial to getting the correct output from the chatbot," Choudhary says. "Before, I would ask regular ChatGPT for the molecular structure notation of ibuprofen, and it would give an incorrect or generic response. With CME, I'll get the right answer to this and many other materials science questions."
The databases, including National Institute of Science and Technology-Joint Automated Repository for Various Integrated Simulations (NIST-JARVIS), the National Institutes of Health-Chemistry Agent Connecting Tool Usage to Science (NIH-CACTUS), and Materials Project, consistently update CME with the most recent materials science findings, he says.
"Materials Explorer is correct because these databases are automatically updated with new papers; it runs itself and pulls from the newest journals," Choudhary says.
To test its resistance to hallucinations, Choudhary compared CME to ChatGPT 4 and ChemCrow, an AI agent geared to solve chemistry-related tasks. From asking the molecular formula for aspirin to interpreting phase diagrams, CME got all eight answers correct, whereas the other models gave only five accurate responses.
Choudhary is now working to develop the platform further by adding advanced materials modeling tools, automated literature reviews, and more. He is also developing an open-source platform which is available at AtomGPT.org. Contrary to the closed-source model of CME, which doesn't enable users to edit the code that Choudhary established, Atom GPT allows select users to change the code and improve its ability to answer materials science questions.
"The ultimate goal is to make ChatGPT Materials Explorer the one-stop research assistant that can help materials scientists with computer simulations, data analysis, and other methods that advance the field," Choudhary says. "What started as a fun project on the weekends has turned into something that could be a useful career resource for materials scientists."
