The Non-AI-pocalypse in Materials Science

By now, everyone has heard about the OpenAI-Navier-Stokes saga and the resulting existential crisis in mathematics. My own experience with the rapidly improving capabilities of AI in materials simulation has pushed me to ask the same question of my own field: will materials theorists become redundant?

At the risk of spoiling the conclusion, my answer is no, because I have spent my entire career trying to make myself redundant and I haven’t quite succeeded.

2006: My first paper as a PhD student was the Li-Fe-P-O phase diagram. It took the better part of my first year, most of it spent debugging calculations, fixing input files by hand, and queuing multi-day jobs on a congested cluster. The analysis and writing took less than a month.

2011: Fast forward five years. I started pymatgen and custodian. Input generation and error recovery stopped being manual. The Materials Project launched the same year, with pymatgen and custodian as building blocks. Anyone could now pull up the phase diagram of any chemical system in seconds (the Phase Diagram app was the first app I built for MP, other than search).

Even then, the first indications of an existential crisis for materials theorists were already there. Many experimental groups increasingly decided they could run their own calculations. Phase stability and battery voltage predictions no longer required a simulation specialist.

So I moved on. I declined collaborations in which nothing interesting was asked of the theorist and the request amounted to routine DFT calculations, even when a high-profile paper was on offer. I turned my group to problems that were still hard: ion diffusion, luminescence in phosphors, defect properties.

2022: A decade later, universal machine learning interatomic potentials, or foundation potentials, arrived on the scene, with my group contributing our small part to their development. A whole swath of once-hard problems suddenly became easy: long-timescale, large-scale molecular dynamics, phonons, elastic constants. With MatCalc, we tried to make them easier still: a property calculation that once needed a bespoke workflow and days of DFT on a supercomputer became a few lines of Python and minutes on a laptop.

Then LLMs matured. LLMs and foundation potentials, orchestrated via MatCalc, turn out to be an especially lethal combination. Even with a fairly high-level prompt, the LLMs are now able to generate a physically grounded sequence of calculations, write the glue code, and deliver the analysis in a nicely formatted report with graphics. Today, my group members and I can test a PhD's worth of hypotheses this way.

More than ever, it seems like the end of the materials theorist is nigh.

So why do I believe otherwise? My observations of users of these powerful new tools (fellow scientists, postdocs and students) showed me that the human driver of these tools matters more than ever. There is a certain aspect of - for lack of a better word - taste in choosing which questions are important and which hypotheses would probabilistically yield the most useful insights. A poorly posed initial question results in AI scientific slop, an unfortunate side effect that is now becoming far too common.

Indeed, the works that I am personally most proud of from my group are not the highly cited tool papers, but those that leverage these tools to make an unexpected materials discovery (e.g., Zhenbin Wang's discovery and confirmation of the first Sr-Al-Li-O compound and its application as a phosphor) or gain new insights (e.g., Xiang-Guo Li's work on short-range order in the NbTaMoW "high-entropy" alloy and our collaboration with the Asta and Ritchie groups on dislocation motion in this alloy).

The theorist whose value was "I can run the DFT calculation" started becoming redundant in 2011. The theorist whose value is knowing what to calculate now wields the throughput of an entire research group. To paraphrase a popular superhero, with that great power comes even greater responsibility to work on harder problems. That, after all, is what science has always been about for me.

** This post is about AI, but not AI-generated. Any dashes, semi-colons, etc. are my writing tics. AI was used to fix grammar and phrasing.**