An intelligent predictive-inverse design framework for cellular steel beams with integrated robustness, sustainability metrics, and automated CAD generation

Abstract

Cellular steel beam design tightly couples geometry, resistance behavior, and fabrication constraints. Current practice relies on simplified, empirical forward checks and causes trial-and-error iterations when multiple geometric variables and opening configurations are considered. This paper presents an inverse design framework, in which structural performance requirements are prescribed first, and compatible geometries are identified systematically. The framework addresses a multi-criteria inverse design setting with non-unique solutions, discrete section availability, and coupled geometric constraints. The framework is built around a physics-informed surrogate model trained on 14,094 finite-element simulations, which acts as a fast forward-analysis operator. The model predicts the beam resistance from geometric, material, and layout parameters. During training, monotonic relationships between key variables and resistance are enforced, consistent with structural mechanics, to ensure stable and physically plausible predictions during inverse use. Inverse design is formulated as a constrained search over parent (unperforated) sections. For a given target resistance and opening configuration, candidate sections are evaluated by the forward surrogate and filtered using geometric admissibility, code-based limits, and target-matching criteria. Robustness to fabrication tolerances is evaluated through controlled geometric perturbations, and feasible designs are ranked using transparent multi-objective measures. Sustainability indicators based on embodied carbon and circularity scenarios are evaluated on the feasible set only, allowing structural and environmental trade-offs to be examined explicitly. The framework supports informed engineering decision-making when mechanical performance and environmental impact must be considered together. Scenario studies illustrate how feasible solution sets evolve with increasing demand and show that near-target designs can be identified without systematic overdesign.