That’s the premise of a paper written by a team of researchers from the University of California, Santa Barbara; University of New Hampshire; and Northwestern University. “These emergent capabilities are exciting for science because they vastly lower the barrier to effective scientific data mining,” they wrote in “Perspective: Large Language Models in Applied Mechanics.” The article appeared in the October 2023 issue of the ASME Journal of Applied Mechanics.
LLMs such as ChatGPT and PaLM have the ability to perform sophisticated text comprehension as well as produce replies in the form of essays, poems, and other formats.
“Alongside their broader societal impacts, these capabilities carry great promise for the physical sciences, including applied mechanics,” the team reported. The paper summarized recent developments in these models, their possible use in mechanics and adjacent fields, and what they could be used for in applied mechanics in the future.
One issue the team addressed is the degree to which these LLMs are limited in their usefulness. LLMs have been known to produce artifacts called hallucinations, in which the model produces a reply to a prompt that is linguistically correct but is also nonsensical or erroneous. For instance, lawyers have been caught using LLMs to help write their briefs due to the inclusion of citations to imaginary court cases. Other problems with the output of the models include multi-state reasoning errors, interpretability, and lack of originality. What do these limitations mean for the adoption of LLMs to applied mechanics?
Such limitations must be addressed before LLMs can be helpful to mechanical engineers and physical scientists, the authors stated. “This tendency threatens any use of LLMs as knowledge bases, and cuts into their advantage over more traditional information retrieval engines,” they wrote.
Proponents of language models within applied mechanics have devised a way to mitigate this issue, however. Researchers suggested that retrieval-augmented generation first retrieve documents related to a query and then prompt the LLM to generate text based on those documents.
Once a mechanics-specific language model can be trained, applied across its entire training corpus, and used to retrieve relevant documents, a second LLM can then be used to synthesize these results into a coherent response. Researchers are working on similar “workarounds” for multi-step reasoning errors, interpretability, and lack of originality.
Data extraction
There are several key emergent capabilities of LLMs that represent promising directions for their use in applied mechanics. For example, information extraction is the first and perhaps most immediate capability in information extraction, especially those that are generated solely by a prompt (called zero-shot learning) or when the prompt is accompanied by one or more examples of the expected output (called few-shot learning).
“With little or no supervision, very large language models (VLLMs) can perform sophisticated reading comprehension tasks over a large number of documents,” the team reported. This capability alone allows for quick iteration on extraction tasks, allowing scientists to efficiently experiment with what information might be helpful to extract at scale from the literature.
But zero- and few-shot learning is not the only bright spot on the researchers’ list, which also included normalizing representations, assisted programming, knowledge recall, and hypothesis generation.
The authors of the paper in the ASME Journal of Applied Mechanics were Neal R. Brodnik, Caelin Muir, McLean P. Echlin, Tresa M. Pollock, and Samantha Daly of the University of California Santa Barbara; Samuel Carton and Satanu Ghosh of the University of New Hampshire; and Doug Downey of Northwestern University.
Looking to the future, the researchers said LLMs have capabilities that scientists can apply to the field of mechanics. Industry expects to see LLMs get better as researchers and developers make the technology not only efficient but scalable. Yet, challenges remain.
The ability of VLLMs to extract information, for example, offers “unprecedented ability to perform complex reading comprehension tasks at scale, higher impact results may arise from using VLLMs for hypothesis generation, or otherwise.”
However, one of the biggest open questions that still remains is when applying LLMs to mechanics, which of these capabilities is likely to produce the most effective results in terms of the overall goals of the field? And that answer will decide the role of LLMs in applied mechanics going forward.
“These answers will remain unknown until mechanics-based LLMs are actually built and applied,” the authors concluded.
Cathy Cecere is membership content program manager.
