How Many Hidden Variables Does a Buckling Steel Beam Need?

Steel design codes decide whether a beam can bend all the way to its full plastic strength using a simple rule: each plate in the cross-section must be thick enough compared with its width. Above a slenderness limit, the plate wrinkles first, which is called local buckling. That wrinkling can then set off twisting of the whole member. Because higher-strength steels carry more stress without getting any stiffer, the limit tightens roughly as $\sqrt{E/F_y}$, where $E$ is the elastic stiffness and $F_y$ is the yield stress. These limits come from expensive shell finite-element studies, meaning simulations that model every plate of the section as a thin surface. They are also validated only up to 70 ksi steel and for a small set of shapes. The practical engineering tool, a one-dimensional beam element that treats each cross-section as rigid, cannot represent this local wrinkling at all.

We propose to build a beam element whose cross-section behaviour is a small learned model. Besides the usual beam strains, it carries a strictly limited number $k$ of hidden "distortion" variables that remember how the section has deformed so far. We then search for the smallest $k$ that reproduces shell-level results. The central bet is that this minimal $k$ is a physical measurement and not only a tuning knob. It should be zero where a section is compact, since rigid-section theory is then enough. It should rise sharply across the compactness boundary, where local and global buckling interact. Mapping that jump across yield strength, orientation, and taper is a data-driven way to find compactness limits in the regimes the codes leave unverified: steels above 70 ksi, unequal-leg angles, and tapered girders.

Core Claim

Two claims are made, and they can be tested separately. The first is a modelling claim. A history-dependent neural model of the cross-section, embedded at each integration point of a one-dimensional beam and allowed a few distortion variables, can reproduce the moment–rotation curves from shell finite-element analysis where local and global buckling are coupled. The target accuracy is enough to recover rotation capacity $R=(\Theta_2-\Theta_1)/\Theta_1$. Here $\Theta_1$ is the elastic rotation at the plastic moment $M_p$, and $\Theta_2$ is the rotation at which capacity falls back below $M_p$. Plastic design requires $R\ge 3$.

The second is a scientific claim. Define normalized slenderness $\bar\lambda=(b/t)/\sqrt{E/F_y}$, the ratio a code limit such as $\lambda_p=0.68\sqrt{E/F_y}$ is written in. The smallest budget $k^*$ that meets a fixed accuracy should switch from zero to positive at a threshold $\bar\lambda^*$. We then check two things. The threshold should agree with published limits in the validated range. It should also stay independent of $F_y$ there, which is what $\sqrt{E/F_y}$ scaling asserts. Our hypothesis is that this independence breaks above 70 ksi, because steels with a low ratio of yield to ultimate strength ($F_y/F_u$) have short strain-hardening plateaus. If $k^*$ does not track the known boundaries in the validated range, the second claim fails.

Prior Work and Positioning

The mechanics results that motivate this work come from Earls and collaborators. Practical Compactness and Bracing Provisions for the Design of Single Angle Beams ran ABAQUS shell models of equal-leg angles in four bending orientations with steels from 50 ksi to HSLA-80. It found compactness limits of 14 for Case 1 and 13 for Case 3 at 50 ksi. It also identified Case 3, where the outstanding leg is in compression, as the orientation where local and global buckling are strongly coupled. Proposed Equal Leg Single Angle Flexural Design Provisions turned this into $\lambda_p=0.68\sqrt{E/F_y}$, but explicitly caps validity at $F_y\le 70$ ksi and $L_b/r_z\le 200$. Yielding Limit State of Tee Stems in Flexural Compression found $h/t_w\approx 0.42\sqrt{E/F_y}$ for tee stems but ignored lateral-torsional interaction. Design of Single Angles Bent About the Major Principal Axis mapped a three-way interaction among $F_y$, $b/t$, and unbraced length.

On Moment Capacity and Flexural Ductility in Doubly Symmetric Web-Tapered I-Girders showed that tapered girders need "ultra-compact" flanges to reach $R\ge 3$: $b_f/2t_f\le 3.0$ for 6-inch flanges and about 6.4 for 8-inch flanges. It also showed that decoupled code checks can be slightly unconservative. These papers supply our ground truth and our validation targets. What they do not supply is a reusable model: each result is a regression fitted to one family under uniform moment.

The method we build on is Physics-augmented neural networks for constitutive modeling of hyperelastic geometrically exact beams. It replaces the costly approach of solving a 2D cross-section problem at every beam integration point with a network that conserves energy and respects symmetry, and it captures warping. It is limited to hyperelastic materials with no path dependence and to point-symmetric sections. Our model differs in exactly the ways the steel problem requires. It has internal state, so it can represent plastic flow, residual stress, and unloading. Its inputs include asymmetric and tapered geometry. Its state budget is explicit and minimized.

The search machinery comes from memory-constrained architecture search. Single-Path Mobile AutoML nests smaller kernels inside a larger "superkernel" so that one training run covers every size. We nest distortion-variable widths the same way, so one run yields models for every $k$. ProxylessNAS supplies the principle of searching directly on the target task under a hardware cost rather than on a proxy. Progressive Differentiable Architecture Search is a cautionary result: architectures chosen in a shallow proxy setting behave differently in the deep target setting. Our analogue is the risk that a $k$ chosen on single-member tests fails inside a full frame analysis, and we test for this directly. The memory that matters for us is not GPU memory during search. It is the state stored at each integration point of the structural solver, which grows with the number of points, $k$, and the loading history. Auto-PINN shows that searching physics networks becomes cheap once the search space is separated into independent choices. We follow that advice and do not claim a new search algorithm.

Research Program

The first step is data. We rebuild the Earls shell-model protocol, including residual stresses from cooling and geometric imperfections. We extend it to three families: equal- and unequal-leg angles in all four orientations, tees, and doubly and singly symmetric tapered I-girders. Yield strength ranges from 50 to 100 ksi. Above 70 ksi we use material curves with reduced $F_y/F_u$ and shortened plateaus. Loading covers uniform moment, linear moment gradients, and cyclic histories.

The second step is extracting training data at the section level. Along each shell model we reduce the kinematics to beam strains (axial, two curvatures, twist, warping) and to the resultant forces. The section model, a small recurrent network with $k$ distortion variables, maps the history of beam strains to resultant forces.

Training uses this objective:

$$\mathcal{L}=\mathbb{E}\big[\|\hat{\mathbf{s}}-\mathbf{s}\|^2\big]+\mu\,\mathbb{E}\big[\,\|\mathbf{h}\|_0\,\big]\quad\text{s.t.}\quad k\le k_{\max}.$$

Here $\hat{\mathbf{s}}$ and $\mathbf{s}$ are the predicted and shell-derived section resultant forces, and $\mathbf{h}$ is the vector of distortion variables. The second term, with weight $\mu$, penalizes active distortion variables through nested gating, so that variable $j$ can be used only if variables $1,\dots,j-1$ are used. The model is also required to dissipate energy rather than create it, so that plastic flow cannot run backwards.

The third step places the trained section model inside a geometrically nonlinear one-dimensional beam solver. Each member is re-analyzed as a beam, and we compare its moment–rotation curve, peak moment $M_n$, and rotation capacity $R$ against the shell result. The fourth step maps $k^*$ over $(\bar\lambda, F_y, \text{orientation}, \text{taper}, L_b)$ and extracts the thresholds $\bar\lambda^*$.

Baselines

The first reference point is the current code approach: decoupled AISC checks and Earls-type closed-form limits. For tapered girders, Miller and Earls report ratios of shell-predicted to AISC-predicted ultimate moment between 0.949 and 1.29 (On Moment Capacity and Flexural Ductility in Doubly Symmetric Web-Tapered I-Girders). That spread is the error band our beam element must beat on the same geometries. Beating it means predicting $M_n$ within ±3% and never on the unsafe side in the coupled cases.

The second reference point is a standard fiber beam element with rigid sections and $J_2$ plasticity, where the yield function uses the von Mises criterion. This is the $k=0$ model without any learning. It is the natural competitor because it is what practitioners actually run. It serves both as the naive reference and as the ablation that isolates distortion memory. It runs on identical meshes and loads.

The third is a hyperelastic section surrogate in the style of Physics-augmented neural networks for constitutive modeling of hyperelastic geometrically exact beams. It gets the same inputs and the same training data, and it is tuned with the same number of hyperparameter trials as our model. We found no reported result on Lacuna for any surrogate on inelastic coupled buckling of asymmetric sections, so we will establish this number ourselves. We expect it to fail on unloading and cyclic curves, and beating it means lower error on post-peak and reversed branches.

The fourth is an unconstrained recurrent section model with a large state, $k=32$. It tests whether our budget costs accuracy. We count a win if the $k^*$ model is within 10% of its error while using several times less solver state.

The fifth is an exhaustive grid over $k$ from 0 to 8, retrained from scratch for each value. It checks that nested gating finds the same $k^*$.

Why not just…?

Why not just regress $M_n$ and $R$ directly from geometry and $F_y$ with a Gaussian process or gradient boosting? For locating $\bar\lambda^*$ under uniform moment, this is a genuine threat and may be as good or better. We will run it on the same shell data and report it honestly. What it cannot do is transfer to loading it was not trained on. A member-level regression trained on uniform moment has no mechanism for predicting behaviour under a gradient or inside a frame. The decisive test is to train both methods on uniform moment and evaluate them on moment gradients and two-member subassemblies. If the regression matches us there, the element's value reduces to frame-level simulation speed, and we will say so.

Why not just run shell finite-element models for everything? For single members that is possible. It is what Earls did, and it remains our ground truth. Frame-level plastic analysis with hundreds of members, many bracing layouts, and cyclic loading is where shell models become impractical and where designers fall back on rigid-section elements. We will report wall-clock time and solver memory per member for shell models and for our element.

Why not just grid-search $k$ and skip the architecture search? For a single section family, a grid works, and we include it as a baseline. Nested gating is only a bookkeeping device that yields every $k$ from one training run across hundreds of families. We claim no novelty for it. If the grid and gating disagree on $k^*$, the grid result takes precedence.

Why not just use a big unconstrained network? The state stored at each integration point grows with the number of points and the length of the loading history. More importantly, an unconstrained state destroys the diagnostic: $k^*$ carries meaning only because it is the minimum needed.

Decisive Checks

Three results would settle the proposal. First, in the validated range, $k^*$ must switch on near the published limits: 13–14 for Case 3 and Case 1 angles at 50 ksi, $0.42\sqrt{E/F_y}$ for tee stems, and the 3.0 versus 6.4 flange limits for 6-inch and 8-inch tapered girders. It must also switch on at the same $\bar\lambda^*$ for 50 and 70 ksi steel. If it misses these, the second claim fails. Second, the element must predict $R$ within ±0.5 on held-out Case 3 and tapered members, including under moment gradients it never saw. Third, a $k^*$ chosen from single-member tests must still meet the accuracy target inside frame subassemblies; if it does not, the analogue of the depth gap is real. If all three hold, we report $\bar\lambda^*(F_y)$ above 70 ksi as proposed compactness limits for high-strength steel, to be confirmed by targeted shell runs.