A graph-based equation search recovers readable constitutive laws for steel, lithium and rubber

The usual way to encode a candidate equation for a machine search is a tree, and the Ningbo team dropped it for a directed graph ^[4]. Nodes carry the operators and the physical variables. Edges carry the computational connections, and the edges can also hold fixed physical constants and tunable material-dependent parameters ^[5]. Putting the parameters on the edges is what lets one search do both jobs at once: it looks for a functional form while fitting the numbers inside that form, so a structure found once can be reused across materials that need different parameter values ^[6].

Yuntian Chen, an EIT associate professor and co-corresponding author, said the encoding was the bottleneck. "By representing mathematical equations as graph structures, we enable simultaneous optimization of equation topology and material parameterization, which resolves this key bottleneck," he said ^[13].

Across the three material systems, the published account does not include error figures for any of the comparisons ^[19]. An engineer would want those, because the question about a discovered law is how it behaves at temperatures and strain rates outside the ones that were fitted. Lithium metal is the hard case there. The release describes its mechanical behaviour as highly sensitive to temperature and strain-rate variation, and difficult to model with traditional methods ^[9].

The interpretability claim is the substantive one ^[3]. An explicit formula can be differentiated, inspected at its limits and implemented by someone who never saw the experimental data; the weights of a fitted network cannot be read that way. The authors describe the output as physically consistent, compact and human-interpretable ^[7]. Hao Xu, the study's lead author, said the older route has a ceiling built into it: "Although this paradigm has achieved great success in mechanics research, predetermined equation structures inherently restrict the model's descriptive and predictive capability" ^[12].

I would expect an established model such as Johnson-Cook to hold its place in production work until other groups recalibrate the discovered equations on their own specimens and report where they fail. Dongxiao Zhang, a chair professor at EIT and corresponding author of the study, said existing forms leave a gap: "Many complex material mechanical behaviors cannot be well described by existing empirical constitutive models" ^[15].

The team says the framework has potential beyond computational mechanics and material modelling, in disciplines that want interpretable physics ^[17].