Pretrained geometry-physics-language surrogate for text-conditioned bridge response screening

Le résumé fourni par la source

At the preliminary bridge design stage, engineers compare candidate schemes under many load cases, where the dominant cost is finite element (FE) model setup rather than the linear-static solve. This study asks whether natural-language load descriptions can serve as a conditioning interface for preliminary response screening under a fixed cable-stayed support topology. On a code-range-calibrated, fully crossed geometry–load dataset (8874 geometries seven load types, 62,118 samples), a staged geometry physics language surrogate with a material-aware, element-connectivity-aware backbone injects load semantics into a pretrained Transformer through adaptive layer normalization. The text interface is viable for screening, reaching an in-distribution displacement nRMSE of 7.73% and an overall von Mises stress of 0.952. Under a matched-protocol, same-information structured baseline it is modestly less accurate (about 1.2 pp), so language is a usability and open-vocabulary contribution rather than an accuracy gain. Adding explicit element connectivity raises the previously weak tower stress from 0.16 to 0.61 in a controlled Stage-2 ablation—a param-matched random-edge control recovers only about 2% of this gain, confirming connectivity rather than capacity—and to a moderate 0.71 on the standard text model; material/section properties are now a model input. On the standard model, shuffling the text and load-vector inputs raises stress error by + 103.5% and + 264.9% (multiple seeds), indicating both conditioning channels are genuinely read. The scope is preliminary screening—global displacement and coarse component-level stress—not design-grade local stress or a replacement for high-fidelity FEA.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.