Detailed analysis is slow, and you run it again and again
A serious engineering analysis is slow to run, and design is iterative by
nature: you change a dimension, a load, or a detail, and run it again. The analyses that
decide a design tend to be the slow ones:
a nonlinear collapse or pushover analysis
a transient time-history run for seismic or blast
a CFD simulation of flow or heat transfer
a detailed connection model with contact and yielding
These run for minutes to hours, sometimes overnight, and you rarely run one
once. The iteration is the bottleneck.
A surrogate is the way around it: "an approximate mathematical model of the
outcome used instead" of the expensive analysis, fit to many real solves. It predicts what
the full run would say in a small fraction of the time, so you can iterate
toward a design in real time and fall back to the real analysis when it counts.
Source: Surrogate model.
What we did
We took one specific problem: the stress in a steel gusset plate connection. We built
the real finite-element model for it (gmsh meshing, quadratic elements, a real sparse
solve), measured what one solve costs, generated thousands of solves across the plate
geometry, and trained a surrogate on them. The numbers below are all measured on this
model.
...
to run our FEM once (measured mean)
...
degrees of freedom in that solve (mean)
...
faster than one FEM run, per surrogate prediction
...
the surrogate's accuracy on held-out tests (R2)
One honest caveat. Our gusset solves in about a second, not the
minutes-to-days that make a surrogate essential in practice. This is an exposition of the
idea on a problem small enough to run the real solver live and check the surrogate against
it. The same approach pays off most where the full analysis is genuinely slow.
Scope, honestly. Linear elastic plane stress,
one plate configuration family, no buckling, no bolt holes. The capacity ratio shown is a
screening number, not a connection design. No language model is involved anywhere in this
demo.