Structural surrogates learn modal quantities instead of curves: tower displacement error 0.73–0.76%, beam 0.029–0.116%
Synopsis
Neural surrogates for structural dynamic response no longer fit the whole frequency-response curve directly; they first predict frequency-independent modal quantities and then reconstruct the response by analytical modal superposition. Across 20,000 designs each for a steel truss tower and a tapered beam, with supplied modal frequencies, 5% reference damping and a declared band, the modal configurations give mean displacement-history errors of 0.73–0.76% and 0.029–0.116% respectively. These figures cover fixed topology and conditional inference, and do not yet establish design-only deployment or an earthquake-engineering assessment.
Interpretation
Structural dynamic surrogates now learn modal quantities and rebuild the frequency response analytically, so the resonant denominators come from physics rather than from a freely fitted network. Earlier direct spectral predictors mapped design variables and query frequency straight to the real and imaginary parts of the response, leaving the network to reproduce each resonance itself. On the 80 m steel truss tower, the direct spectral SIREN configuration gives a mean displacement-history error of 4.54%, while the modal configurations give 0.73–0.76%; the two groups do not share parameter counts or training settings.
The network emits only frequency-independent modal numerators, frequencies, damping and a static remainder, and an analytical decoder supplies the frequency dependence. Modal Rational is that neural map plus analytical response reconstruction; PGMON adds bounded corrections to frequency, damping and residual flexibility, and MOSAIC adds a bounded spectral discrepancy on top. The tower retains 20 frequency-ordered modes over 0–7 Hz with five X-direction response channels; the tapered beam retains nine participation-selected modes over 0–100 Hz, evaluated on 10 test designs times 7 prescribed acceleration histories.
Perspective
The evaluation covers two fixed-topology populations: an 80 m steel truss tower with 32 area factors and 153 mass coordinates over 429 free translational degrees of freedom, and a 4 m tapered box beam with ten station masses, 20,000 designs each and a test set of 10 designs times 7 prescribed acceleration histories. The modal configurations receive dataset-supplied modal frequencies at inference, 5% reference damping is assigned to every mode of both structures, and the errors are reconstruction errors within the declared band.
The reported results correspond to a conditional modal-information budget and are not evidence of design-only accuracy; the tower and beam differ in excitation direction, modal ordering and training settings, so their gap cannot be attributed to mesh size or bandwidth. The corrections do not dominate every metric, tower displacement maxima are several times the means, and each structure has only one training seed.