“ChatGPT is a very impressive tool,” said paper author Zijian Hong, professor at the School of Materials Science and Engineering, Zhejiang University, China. “As a computational materials scientist, I’m always eager to embrace new tools, in particular, new tools in computer science and AI. Since the [birth] of the new ChatGPT, I’m just wondering whether such a tool can assist us in computational materials science.”

Hong explained that for a computational materials task, there are three main steps: building a model or a structure, writing codes for specific scientific software, and preparing data visualization scripts. To test the capability of ChatGPT, he examined it from these aspects.

“ChatGPT can help us prepare scripts to build atomic structure, i.e., the cif file, scripts for running a DFT calculation, and scripts for data visualization,” Hong said. “At least it is trying to help us from chat, although the scripts are not working at all when I accessed on Feb. 20, 2023.”

“But what surprised me is the ability to evolve and learn from communications,” Hong added, “When I accessed 20 days later, it gives me different answers, towards the correct answer. And if I give more hints, such as the correct lattice structure, it can correct by itself, just like a human being.”

“It is still not perfect for sure. For example, it still makes simple mistakes, the consistency of the output is not guaranteed, and the ethical concerns are still there.” Hong said. “But change is really [around] the corner, for computational materials science. We should embrace it rather than avoid it.”

This work is supported by the Fundamental Research Funds for the Central Universities and a startup fund from Zhejiang University.

- This press release was provided by Beijing Institute of Technology Press