Engineers who lighten a bracket, drone arm or vibrating panel have to meet several physical limits at once. Peak stress must stay below a limit, deflection at a key point must stay small, and the lowest natural vibration frequency must stay above the band where the structure is excited. Generative models trained on simulated designs can now propose structures in seconds instead of hours. These include diffusion models, which learn to turn random noise into samples by gradually removing it, and flow-matching models, which learn a smooth path from noise to data. But they learn from datasets of designs that were optimized for a single, loose goal, usually stiffness at a fixed amount of material. The designs that matter in practice sit at tight, multi-constraint bounds, and almost no training data lives there. Asked for them, the models extrapolate and fail.
We propose to grow the training data toward the hard region in steps, using the physics solver as a judge. A generator conditioned on constraint bounds starts where existing data already satisfies all limits. At each round it proposes designs for slightly tighter bounds. Each proposal is checked with a cheap finite-element simulation, the passing designs are kept, the model is retrained on them, and the bounds tighten again. The pace of tightening is set by how often the model's proposals currently pass. The question we test is whether this easy-to-hard ordering, rather than just more filtered data or stronger guidance at sampling time, is what lets a learned designer reach tight multi-constraint designs on a fixed simulation budget.
Core Claim
For amortized multi-constraint design, meaning one trained model serving many load cases and boundary conditions, we claim that the binding bottleneck is data coverage near the feasibility boundary, not model capacity or guidance strength. We further claim that a self-paced curriculum over constraint severity, checked by finite-element simulation, reaches tight bounds with markedly higher verified feasibility than generating directly at the target bounds under the same budget of simulation calls. The new ingredient is the axis of the curriculum. Difficulty is defined by how tight the stress, displacement, frequency and volume bounds are, and progress is gated by verified feasibility. Prior curricula used question format or a scalar PDE coefficient instead.
Let $c = (v, \sigma_{\max}, \delta_{\max}, f_{\min})$ be the target volume fraction, stress limit, displacement limit and frequency floor, all normalized per boundary condition so that $c = 0$ marks the loosest level the seed data satisfies and $c = 1$ marks the deployment target. At curriculum level $c_k$ the generator proposes designs, and the finite-element check keeps the subset $S_k$ that satisfies every constraint. The level then advances only if the feasibility rate clears a threshold $\tau$:
$$c_{k+1} = c_k + \eta \, \mathbb{1}\left[\,|S_k| / N \ge \tau\,\right] \, u_k,$$
Here $N$ is the number of proposals per round, $\eta$ is the step size, and $u_k$ is a unit direction choosing which constraint to tighten. By default $u_k$ tightens the constraint with the highest current pass rate. This gate plays the role of the pace parameter in self-paced learning, but it is set by physical feasibility rather than by training loss.
How This Sits Against Prior Work
Two results motivate the proposal. Characterizing possible failure modes in physics-informed neural networks shows that neural PDE solvers fail because stiff physics terms make training ill-conditioned, not because the networks lack capacity. Starting from a low convection coefficient and warm-starting upward cut error from nearly 100% to roughly 2%. That paper uses a curriculum over one scalar coefficient while fitting a single solution. We use a curriculum over a vector of inequality bounds while training a generator that serves many problems. A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula proves finite-sample advantages for filtering and ordering tasks by difficulty during self-improvement. We supply a concrete structural instance in which the filter is an exact physics verifier, and we test the theory's ordering prediction directly.
The method borrows from three sources. Self-Paced Learning for Latent Variable Models gives the idea of an annealed inclusion threshold. On motif finding it got stuck in bad local minima once in 100 runs, against 18 for the uniform alternative. We replace its loss-based easiness with constraint satisfaction measured by simulation. Diffusion Models Beat GANs on Topology Optimization (TopoDiff) and Minimizing Structural Vibrations via Guided Flow Matching Design Optimization supply the generators and the practice of steering samples with gradients from learned physics predictors. Both train once, uniformly, on a fixed offline set, and both target one objective: stiffness in TopoDiff and vibration in the flow-matching work. We keep their architectures and change what data they see and in what order.
Other work constrains the design. Physics-consistent deep learning for structural topology optimization finds that one-shot prediction of the final design is much worse than models that follow the optimizer's sequence, with 5.76x higher compliance error in 2D. We therefore keep an iterative generator. TopoDiff's own out-of-distribution results show that single-pass GANs produce infeasible designs about 68% of the time, which rules out a GAN backbone. Self-directed online machine learning for topology optimization (SOLO) is the closest relative in spending simulation calls online. However, it optimizes one problem at a time and samples by proximity to the current predicted optimum, with no schedule over constraint severity. CReFT-CAD shows that staged curricula prevent policy collapse in engineering reasoning, but it stages question formats rather than physics.
Research Program
The main testbed extends TopoDiff's 2D 64×64 setting. We keep its distribution of randomized loads and supports and its two test levels: Level 1 holds unseen combinations of known boundary conditions, and Level 2 holds new boundary conditions. The objective becomes minimum volume, subject to a global stress limit (computed with a smooth p-norm aggregate of von Mises stress), a displacement limit at the load point, and a first-eigenfrequency floor. The 2D linear-elastic solve plus the eigenvalue solve takes a fraction of a second, so the verification loop is affordable.
The seed set is TopoDiff's 33,000 compliance-optimized SIMP designs (SIMP is the standard density-based method). Each design is re-evaluated for stress, displacement and frequency, which labels the constraint levels it already satisfies. Target bounds are set per boundary condition to be tighter than any seed design for that case. This makes the deployment target out of distribution by construction.
The generator is TopoDiff's conditional diffusion model with the bound vector $c$ added to its conditioning. At sampling time it is guided by small learned predictors of stress, displacement and frequency, plus TopoDiff's floating-material classifier. These predictors are retrained each round on all designs verified so far, including failures. Failures matter because they show the predictors exactly where the feasibility boundary lies. We will study how the gate threshold $\tau$, the step $\eta$ and the tightening direction $u_k$ affect results. We will also report which constraint orderings help. For example, it is unknown whether tightening the frequency floor before stress makes later stress tightening easier.
A second testbed checks that the effect is not specific to compliance-style layouts. It uses the van Delden vibrating-plate setting from Minimizing Structural Vibrations via Guided Diffusion Design Optimization and its flow-matching successor. We keep their target of minimizing mean squared velocity in the 100–200 Hz band and add two constraints that those papers lack: a first-eigenfrequency floor and a cap on beaded area.
Baselines
The naive reference is the generator trained once on the seed set and conditioned directly on the target bounds, with no extra simulation calls. This measures how badly uniform training extrapolates to tight multi-constraint targets. The comparison is fair because it uses the same architecture and seed data. Every method should beat it by a wide margin.
The strongest existing learned method is TopoDiff-style guided sampling at the target bounds, with all four physics predictors summed into the guidance and the guidance weights tuned by grid search. Its predictors are trained on the seed set plus extra designs evaluated by finite-element simulation, using the same total budget of simulation calls as our curriculum. TopoDiff reports a feasibility failure rate of about 6% on its single-objective out-of-distribution test, an eightfold reduction in compliance error over GANs, and 21.59 seconds per design. The flow-matching plate model reports a 9% (2.4 dB) reduction in mean vibration over genetic optimization and eight designs in about 11 seconds. Neither number transfers to multi-constraint feasibility. We found no reported result on Lacuna for verified multi-constraint feasibility of a generative structural model. Beating this baseline means at least a 2x higher verified all-constraints feasibility rate on Level 2 at the target bounds, under equal simulation budget and equal tuning trials, with no increase in floating material.
The ablation that isolates ordering is filtered self-training without a curriculum. It runs the same propose, verify, keep and retrain loop, but every round proposes at target-level bounds, or at bounds drawn uniformly between seed and target. It uses the same number of rounds and the same simulation budget. If this matches the curriculum, the gain comes from verifier filtering, not from ordering, and our core claim fails. A second ablation replaces the feasibility gate with a fixed linear schedule, which tests whether the self-paced pacing earns its place.
The per-instance gold standard is SIMP with stress, displacement and frequency constraints, using continuation on the penalty exponent, run separately for each test boundary condition. It defines the best attainable volume. We report its cost in simulation calls per design alongside our amortized cost, which covers training calls plus per-design verification. We also run SOLO on a 50-instance subset as the online-surrogate competitor. SOLO reports reducing simulation calls from 200,000 to 501 on a compliance problem, and we compare against it on calls per solved instance.
Why not just…?
Why not just use heavier multi-predictor guidance at the target bounds? This is the most likely threat. TopoDiff's guidance already cuts floating material elevenfold. The issue is that the guidance predictors themselves are trained where the data is, and the second-tier problem documented for TopoDiff is exactly that predictors give unreliable gradients out of distribution. The equal-budget guided baseline above settles this. If it reaches comparable feasibility, the curriculum is unnecessary, and we will report that.
Why not just spend the simulation budget generating constrained SIMP data at the target bounds? This is a real alternative. Each constrained SIMP run costs hundreds of simulation calls, whereas one verification costs a single call. We will run an equal-budget variant that trains on as many target-level constrained SIMP designs as the budget allows. The curriculum must beat it on Level 2 feasibility or on volume at matched feasibility to justify itself.
Why not just optimize each design with constrained SIMP and continuation? For a one-off design this is the right tool. Classical practice already ramps the penalty exponent, and we use it as the gold standard rather than a competitor. The proposal targets the regime where many boundary conditions must be served quickly. There, SIMP's per-design cost, which Physics-consistent deep learning for structural topology optimization reports as 390 seconds for a 3D case, is what amortization removes. We report the break-even number of designs.
Why not just filter without ordering? This is the ordering ablation in the Baselines section, and it is the decisive test of the claim.
Proposed Experiments
On both testbeds the primary metric is the verified feasibility rate, meaning the fraction of generated designs that pass every constraint in an independent finite-element check, on Level 1 and Level 2. Secondary metrics are volume (or vibration level on plates) of feasible designs relative to the per-instance constrained-SIMP reference, floating-material percentage, total simulation calls, and wall-clock time per design. We run five seeds and three simulation budgets so that results can be read as curves of feasibility against budget.
The decisive check is the ordering prediction. As targets are pushed farther beyond the seed data, the gap between the curriculum and the unordered filtered loop should widen. Near the seed data the two should tie, because filtering alone suffices. If the gap does not grow with distance, or if the unordered loop matches the curriculum at every budget, we conclude that verifier filtering, not curriculum, drives the gains.
A second check asks whether the feasibility gate adapts sensibly. It should advance quickly along constraints that are easy for a given boundary condition and stall on binding ones. It should also outperform the best fixed schedule found with the same tuning budget, which would address the schedule-sensitivity problem noted for both self-paced learning and PINN curricula.
The plate testbed tests transfer. Here the curriculum must beat the guided flow-matching baseline once the added frequency floor and area cap are imposed.
References
- Characterizing possible failure modes in physics-informed neural networks
- A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula
- Self-Paced Learning for Latent Variable Models
- Diffusion Models Beat GANs on Topology Optimization
- Minimizing Structural Vibrations via Guided Flow Matching Design Optimization
- Algorithmically-consistent deep learning frameworks for structural topology optimization
- Self-directed online machine learning for topology optimization
- CReFT-CAD: Boosting Orthographic Projection Reasoning for CAD via Reinforcement Fine-Tuning
- Minimizing Structural Vibrations via Guided Diffusion Design Optimization