Learning Where Optimized Steel Members Buckle: A Yield-Normalized Member Surrogate for Thousand-Variable Structural Design

When an optimizer sizes a steel structure, it thins every plate element as far as the constraints allow. If the design uses a higher-strength steel, it thins them further, because each member can now carry more stress. The stiffness of steel, however, does not change with its grade, so those thinner plates wrinkle sooner. In asymmetric members such as the single angles that make up transmission towers, that local wrinkling also triggers sideways twisting of the whole member. Optimizers capable of handling thousands of design variables never see any of this, because they treat members as solid bars that can fail only by yielding or by textbook column buckling. The structural mechanics literature has characterized this failure carefully with detailed finite element studies, but only one member at a time, never inside an optimizer.

This note proposes to close that gap with a single learned model of member strength that the optimizer queries in place of design-code formulas. The model takes a handful of dimensionless inputs, in which the yield strength is folded into the slenderness variables using the known square-root scaling. Two claims make the approach worth testing. First, the coupled local-global failure of a member depends on so few combined variables that it can be learned from a few hundred shell simulations, even when the structure has a thousand members. Second, the optimizer will deliberately push members into exactly the region where local and global buckling occur at the same load, which is the region where existing simple checks are least reliable and where real strength drops below both predictions. If both claims hold, finite element cost scales with the dimensionless member space instead of the structural design space, and the optimized designs remain safe after full verification.

Core Claim

Two lines of work describe the member-level physics. Earls and Galambos (Practical Compactness and Bracing Provisions for the Design of Single Angle Beams) and Earls (Proposed Equal Leg Single Angle Flexural Design Provisions…) used nonlinear shell finite element analysis of single angles to show two things. Compactness, the ability of a plate element to yield fully before it wrinkles, depends on bending orientation, with a slenderness limit $\lambda_p = 13$ when the outstanding leg is compressed and $\lambda_p = 14$ when it is in tension, in 50 ksi steel. That limit also depends on the unbraced length, which is direct evidence of local-global coupling. Earls and Volle (Yielding Limit State of Tee Stems in Flexural Compression) found that tee stems reach the plastic rotation capacity of $R \ge 3$ at $b/t \approx 0.42\sqrt{E/F_y}$, where $E$ is Young's modulus and $F_y$ the yield stress. Miller and Earls (On Moment Capacity and Flexural Ductility in Doubly Symmetric Web-Tapered I-Girders) showed that tapered girders need "ultra-compact" flanges with $b_f/2t_f \le 3.0$ for 6 in flanges. In the same study, code strength predictions ranged from unconservative to conservative, with finite element to code moment ratios of 0.949 to 1.29. These papers motivate the proposal. Each is a forward parametric study on Grade 50 steel at fixed sizes, and none was used to drive design.

The proposal turns these findings into a hypothesis that can be falsified. Define the yield-normalized plate slenderness $\bar\lambda = (b/t)\sqrt{F_y/E}$ and the yield-normalized member slenderness $\bar L = (L_b/r)\sqrt{F_y/E}$, where $b/t$ is the leg width-to-thickness ratio, $L_b$ the unbraced length and $r$ the governing radius of gyration. The hypothesis is that the ultimate capacity ratio and the ductility of a member obey

$$\frac{M_u}{M_p},\; R \;\approx\; g\!\left(\bar\lambda,\ \bar L,\ \text{orientation},\ e/b,\ \beta,\ \eta\right).$$

Here $M_u$ is the ultimate moment, $M_p$ the plastic moment, $e/b$ the connection eccentricity relative to leg width, $\beta$ the end-moment ratio describing the moment gradient, and $\eta$ an optional descriptor of strain hardening, such as the yield-to-tensile ratio. If $g$ needs no $\eta$ term, then $F_y$ is not an independent axis, and a model trained mostly on mild steel transfers to high-strength grades. The file names the failure of this scaling for steels above 690 MPa as an open problem, so whether $\eta$ is required is itself a result.

Research Program

The first component is a member-level shell finite element generator modeled on the Earls protocol. It uses ABAQUS S9R5 shell elements, imperfections seeded from the first ten elastic buckling modes, hot-rolling residual stresses, and elastoplasticity with strain hardening. Equal and unequal-leg angles are loaded as beam-columns through one leg, following the connection eccentricity mapped by Earls and Keelor (Toward the Simplified Design of Single-Angle Beam-Columns), across moment gradients and steel grades from 345 MPa to above 690 MPa. Tees and tapered I-sections are added as secondary families. The generator is accepted only after it reproduces the published limits: $\lambda_p = 13$ and $14$ for the two angle cases, $h/t_w = 10$ for Grade 50 tee stems, and $b_f/2t_f \le 3.0$ for 6 in tapered flanges.

The second component is the surrogate $g$. It is a small monotonic neural network over the dimensionless inputs, constrained so that capacity cannot increase as slenderness increases. Alongside the network we fit a parametric interaction formula of comparable form, so the comparison can show whether a neural model is needed at all. The training points are chosen actively. The approach borrows the self-directed sampling of SOLO (Self-directed online machine learning for topology optimization), which concentrates new evaluations near the optimizer's current best design. We apply that idea at the member level: new shell runs are placed where the structural optimizer's active constraints currently sit in dimensionless space. This is the main methodological difference from SOLO. SOLO learns the structure-level objective, and its surrogate input grows with the number of design variables $d$. Our learned object has fixed input dimension regardless of $d$, and the structure-level forces come from a cheap frame analysis.

The third component is the outer optimizer, which sizes leg widths, thicknesses, and steel grade per member group. We test it with SOLO-style search and with CMA-type search, using CatCMA (CatCMA: Stochastic Optimization for Mixed-Category Problems) for the categorical grade choice. Design dimension runs up to roughly 1,000. Member strength enters as a constraint through $g$, with a margin calibrated from the surrogate's held-out error.

The fourth component is diagnostic. For every optimized design we record how many members land where the local and global critical loads are nearly equal. We then measure, by direct shell analysis, how far below both individual buckling loads their true strength falls. This tests the claim that optimization drives designs into the coupled knock-down zone.

Baselines

The naive reference is the optimizer used in SOLO and AutoTruss (Automatic Truss Design with Reinforcement Learning), which constrains only stress, displacement and textbook column buckling $\pi^2EI/L^2$. Its optimized designs are re-checked with member shell analysis. It will almost certainly look lightest, and the question is how many of its members actually fail. Beating it means producing designs with zero shell-verified member failures at a mass penalty we report honestly.

The strongest practical competitor is the same optimizer with the AISC-style compactness and single-angle provisions as algebraic constraints. This is what a practicing engineer would do, and for prismatic members of mild steel it may be adequate. It runs on the same frame model, outer optimizer and iteration budget. Our method beats it if it either finds lighter shell-verified designs, by exploiting the tension-leg reserve up to shape factor 1.8 that Earls documented, or avoids shell-verified failures that the code constraints admit. Tapered members and high-strength grades are the expected places for this. If code constraints match our method on both mass and safety for prismatic mild-steel towers, we report that and restrict the claim.

The strongest learning baseline is SOLO applied directly at the structure level, with a structure-level neural surrogate and the same member shell checks used as ground truth. It receives the same total budget of shell analyses. We also run the Gaussian-process Bayesian optimization of Perera et al. (Bayesian Buckling Load Optimisation for Structures with Geometric Uncertainties), but only on a reduced problem of fewer than 20 grouped variables, the regime its authors state it is limited to.

Two ablations isolate the key ingredients. In the first, the surrogate is given raw $b/t$, $L_b$ and $F_y$ instead of the yield-normalized variables, which isolates the dimensional collapse. In the second, shell runs are placed by uniform space-filling sampling instead of at the optimizer's active constraints, which isolates the targeted sampling.

Why not just…?

Why not just use the design-code compactness limits as constraints? This is the most serious alternative. The file's evidence against it is specific to certain member types: the 0.949 unconservative ratio for tapered girders, the 1.5 $M_y$ cap that ignores tension-leg reserves of 1.8, and the dependence of $\lambda_p$ on unbraced length in compressed-leg angles. None of this shows the code is inadequate for ordinary prismatic tower angles in Grade 50 steel. The experiments therefore stratify results by member family and by grade. We expect code constraints to break down in high-strength grades and in the coupled region where optimized members cluster. If they do not, the proposal shrinks to a recommendation for tapered and high-strength members only.

Why not just fit a closed-form interaction curve to the Earls data instead of a neural surrogate? If a two- or three-parameter curve in $\bar\lambda$ and $\bar L$ matches the network on held-out shell runs, the simpler curve should be used. That outcome would still be useful, because the contribution lies in the yield-normalized collapse and the active-constraint sampling loop, not in the network. The fitted curve is run as a full alternative surrogate inside the same pipeline.

Why not just run SOLO on the whole structure with shell analysis? A structure-level surrogate must learn a function of $d \approx 1{,}000$ inputs, and each true evaluation would be a full-structure shell analysis taking minutes to hours. SOLO's reported savings, 100 to 1,000 times fewer FEM calls on its 1,008-variable tower, were obtained with linear analyses, and the file contains no evidence that they survive nonlinear post-buckling evaluations. The equal-shell-budget comparison measures this directly. The real threat is that member-level decomposition misses interactions between members, such as moment redistribution after one member buckles. The decisive check below addresses that risk.

Why not just use Gaussian-process Bayesian optimization? Perera et al. state that their method is limited to fewer than 20 design variables. We include it only on the grouped reduced problem, where it serves as a sanity check on solution quality, not as a competitor at full scale.

Proposed Experiments and Decisive Checks

The structural testbeds are the 1,008-variable transmission and antenna tower from SOLO and the 3D Cantilever Sundial from AutoTruss. In both, members are re-specified as single angles connected through one leg, with steel grade added as a design variable. For reference, SOLO reports reaching optimal weight with 100 to 1,000 times fewer FEM calculations than heuristics on the tower. AutoTruss reports layouts 25.1% lighter than prior methods on the Sundial, and an average 6.8% mass reduction over AlphaTruss and SEOIGE on the 2D trusses. Both results use bar-only constraints and are not directly comparable to ours. No reported result was found on Lacuna for optimization of either benchmark under plate-level compactness or coupled local-global buckling constraints, so all baselines in this proposal are run fresh. Member-level validation uses the published angle, tee and tapered-girder limits listed above.

The metrics are mass, the fraction of members that fail shell verification, the minimum shell-verified strength margin, total shell analyses consumed, and the surrogate's held-out error, reported separately inside and outside the coupled region.

Three checks decide whether the central claims hold. The first is transfer across grades. The surrogate is trained only on 345 MPa and 550 MPa grades and tested on shell runs above 690 MPa. If the normalized surrogate without $\eta$ keeps its error within the calibrated margin while the raw-input ablation does not, the yield-normalized collapse is confirmed. If error rises sharply above 690 MPa, that maps where the $\sqrt{E/F_y}$ scaling breaks down. The second is the clustering test. We count the fraction of optimized members with local and global critical loads within 10% of each other under the naive, code-constrained and proposed optimizers, and measure the shell-verified knock-down for those members. The third is the system-level test. Ten optimized designs per method undergo full nonlinear shell analysis of their critical substructures, the most heavily used panels including the connections between adjacent members. If our designs fail at the system level while every member passes individually, member decomposition is insufficient. That would be a clear negative result, and it would locate exactly which coupling between members has to be added.