betaFeedback

WorksComputer Science

Meta-neural Topology Optimization: Knowledge Infusion with Meta-learning

Abstract

When faced with novel design problems, traditional topology optimization methods discard all prior design experience and start from a uniform initial guess. While this avoids biasing the optimizer towards any particular solution, it also means that many computationally expensive iterations are needed to converge. Existing machine learning approaches address this through data-driven design prediction, but require large datasets of pre-optimized structures and often struggle to generalize across boundary conditions and mesh resolutions. We propose a new method, termed meta-neural topology optimization, which uses a meta-learning algorithm to learn effective initial designs for topology optimization with neural field parameterizations – continuous, mesh-independent representations that encode material distributions in the weights of a neural network. Through bilevel optimization, our method distills reusable design knowledge from partial optimization trajectories, eliminating the need for pre-optimized training data. By conditioning the neural field on strain energy fields of reference designs, a single set of learned parameters encodes problem-specific initial structures for diverse boundary conditions. We evaluate our approach on 3000 compliance minimization tasks across in-distribution, out-of-distribution, and cross-resolution scenarios. Our method converges in fewer iterations in 57.6% of in-distribution and 74.1% of cross-resolution tasks, while maintaining design quality competitive with standard density-based topology optimization. Notably, initializations learned on coarse meshes transfer successfully to discretizations four times finer. Code is available at https://github.com/bessagroup/metatopia.

Graphical Abstract

Figure 1. Traditional methods — Start from a blank canvas for each new problem; Discard previous design experience; Require many iterations to converge. Our approach — generic initial designs; meta-learning across diverse training problems; problem-specific initial designs; fewer iterations; many iterations; optimized designs. Key results — Discovery of physics based priors; Consistently accelerated convergence; Successful transfer to finer meshes.

Highlights

Meta-neural Topology Optimization: Knowledge Infusion with Meta-learning

Igor Kuszczak, Gaweł Kuś, Federico Bosi, Miguel A. Bessa

  • Meta-learning infuses past optimization experience into topology optimization.

  • Learned initializations consistently accelerate neural topology optimization.

  • This narrows the gap between standard and neural topology optimization.

  • Initializations learned on coarse meshes transfer to unseen finer meshes.

Introduction

Topology optimization (TO) has transformed computational design across multiple scales and disciplines. Its successful applications span from nanometer-scale optical cavities that transcend traditional limitations on light confinement [1] to full-scale aircraft wing designs with unprecedented structural efficiency [2]. While the field has seen significant advancements since the pioneering work by Bendsøe and Kikuchi [3], one fundamental aspect of human design capability remains underexplored: learning from experience.

Traditional TO approaches adopt a tabula rasa strategy, reinitializing the design variables before every run. This practice persists despite substantial evidence that strategic initial guesses can accelerate convergence [4]. Since each TO iteration requires at least one computationally expensive finite element analysis, and many iterations are needed to reach optimized designs, the lack of knowledge reuse represents a significant missed opportunity.

To develop TO methods that learn from experience, we must first understand the limitations of knowledge transfer in standard approaches. In density-based methods [5], the design domain is discretized into finite elements, and each element is assigned a density value between 0 and 1, indicating the material occupancy. The density distribution is then optimized for the selected objective using a gradient-based optimizer, commonly the Method of Moving Asymptotes (MMA) [6]. Since the design variables are defined on the computational mesh, they must be redefined for each new geometry and mesh resolution [7], preventing the direct transfer of learned representations.

Neural topology optimization addresses this limitation by encoding the density distribution in the learnable parameters of a neural network. Approaches based on neural fields [8] provide continuous, mesh-agnostic representations by mapping spatial coordinates x\mathbf{x} to density values ρ(x;θ)=fθ(x)\boldsymbol{\rho}(\mathbf{x}; \boldsymbol{\theta}) = f_{\boldsymbol{\theta}}(\mathbf{x}), enabling parameters to transfer across different geometries and resolutions. Neural TO has further been shown to reshape the optimization landscape, uncovering designs that mesh-based parameterizations cannot access [9]. However, this reshaping cuts both ways: neural parameterizations can create highly non-convex landscapes that significantly hinder convergence, even in simple compliance minimization problems [10], rendering the quality of the initial parameters critical.

Together, these properties create both the opportunity and the need for learned initializations in neural TO. Prior neural TO work has relied on heuristic initializations such as a uniform-density solution [11, 12] or transfer learning from a small set of pre-solved reference designs [9]. Unlike transfer learning, meta-learning is designed to distill generalizable knowledge from a distribution of related tasks [13], directly optimizing the learning process to accelerate adaptation to unseen problems. Gradient-based methods such as MAML [14] and Reptile [15] have been applied to neural fields across multiple modalities [16, 17, 18] and engineering problems [19, 20, 21], but meta-learning has not been explored in the context of topology optimization.

In this work, we introduce meta-neural topology optimization, which applies meta-learning to neural TO to distill a reusable network initialization from a distribution of related design tasks [13]. As surveyed in § 4, our approach is fundamentally different from existing machine-learning-based TO methods, which require precomputed datasets of optimized designs. Our contributions are:

  • Meta-learning for neural TO. We present the first application of meta-learning to neural topology optimization, learning a reusable initialization from a distribution of related tasks with no precomputed designs.

  • Faster convergence and cross-resolution transfer. Meta-learned initializations consistently reduce iterations to convergence and transfer from lower- to higher-resolution problems.

  • Emergent physical prior. Our approach rediscovers strain-energy initialization as a physically meaningful prior that also accelerates the standard density-based optimizer.

Related Work

Most machine-learning approaches to topology optimization fall into two categories: direct-design methods that predict optimized structures in a single forward pass, and acceleration methods that skip portions of a classical optimizer. Direct-design approaches span CNN prediction from coarse-resolution inputs [22], conditional GANs with physical-field conditioning [23], neural-field-based approaches with post-processing via brief SIMP optimization [24], and latent space diffusion [25, 26]; however, the predicted topology cannot alter which local optimum the downstream physics solver reaches. Acceleration methods take various forms, including predicting final designs from partial SIMP runs [27], mapping early-iteration density histories to near-converged states [28], and upsampling coarse-mesh solutions via super-resolution [29]. Both categories achieve fast design synthesis at inference time and, in recent instances, near-SIMP compliance quality within their training distributions [24]. Despite their methodological diversity, every approach requires a precomputed dataset of optimized designs generated by a reference optimizer, inheriting its biases and local-minima structure. The surveys by Woldseth et al. [30] and Shin et al. [31] provide comprehensive coverage of both categories.

A separate line of work uses neural networks not to predict designs from a training set, but as the representation of the density field itself [32, 33, 12, 11]. Network weights are the design variables, optimized directly against the physical objective, so the full topology optimization pipeline is retained and no precomputed training data is required. This approach offers continuous, mesh-agnostic representations and freedom from checkerboard artifacts [34, 33], but also reshapes the optimization landscape in ways that are not always beneficial. Initialization is particularly consequential in this setting: Sanu et al. [10] showed that neural parameterizations can produce highly non-convex compliance landscapes, while Herrmann et al. [9] showed that a well-chosen initialization can reach local optima inaccessible to standard methods. Specifically, Herrmann et al. [9] pretrained a neural parameterization based on the U-Net architecture on a small collection of designs conventionally optimized across a range of acoustic frequencies, and found that this transfer-learned initialization consistently outperforms random initialization on acoustic TO. While the approach was effective, it was limited to a single set of boundary conditions, and still relied on a small dataset of conventionally-optimized structures.

Methods

Topology optimization

We focus on compliance minimization, where the goal is to distribute the material within a fixed design domain such that the global stiffness of the resulting structure is maximized. We consider a typical formulation, where the problem is discretized with finite elements [4]:

minimize⁡ρc(ρ)=fTu(ρ)(1)\underset{\rho}{\operatorname{minimize}}\quad c(\rho) = \mathbf{f}^{\mathsf{T}}\mathbf{u}(\rho) \tag*{(1)}
K(ρ)u=f(2)\mathbf{K}(\rho)\mathbf{u} = \mathbf{f} \tag*{(2)}
∑e=1Nρeve≤V∗(3)\sum_{e=1}^{N} \rho_e v_e \le V^{*} \tag*{(3)}
0≤ρe≤1;∀e∈{1,…,N}.(4)0 \le\rho_e \le1;\quad\forall e \in\{1,\ldots,N\}. \tag*{(4)}

Here, ρ\rho is a vector of elementwise-constant densities ρe\rho_e. The stiffness matrix, global displacement vector, and force vector are denoted by K\mathbf{K}, u\mathbf{u}, and f\mathbf{f}, respectively. The target volume is represented by V∗V^{*} with vev_e denoting the volumes of individual elements ee.

The global stiffness matrix K(ρ)\mathbf{K}(\rho) is assembled from element stiffness matrices, each scaled by the corresponding elementwise stiffness EeE_e. These are related to densities through the modified Solid Isotropic Material with Penalization interpolation [4]:

Ee(ρe)=Emin⁡+ρep(Emax⁡−Emin⁡),(5)E_e(\rho_e) = E_{\min} + \rho_e^p(E_{\max} - E_{\min}), \tag*{(5)}

where Emax⁡E_{\max} is the Young’s modulus of the base material, Emin⁡E_{\min} is an artificial stiffness value assigned to the void regions, and pp is a scalar penalization factor. In all our experiments, we use the standard values of Emax⁡=1.0E_{\max} = 1.0, Emin⁡=10−9E_{\min} = 10^{-9}, p=3p = 3, and a Poisson’s ratio of ν=0.3\nu= 0.3.

Optimization follows the nested analysis and design approach, where we iteratively (i) solve the discretized equilibrium equation (2) for the displacement field u\mathbf{u}, given a density distribution ρ\rho, (ii) evaluate the objective function (1), and (iii) update the densities considering the remaining constraints. In this approach, displacements u(ρ)\mathbf{u}(\rho) are implicit functions of the densities, satisfying equilibrium at each optimization iteration.

To compute the required gradients, standard topology optimization implementations derive problem-specific adjoint equations for each objective function. In our implementation, we instead apply implicit differentiation to the linear solver [35], which allows us to compute numerically exact gradients for arbitrary objective functions while natively supporting higher-order derivatives.

Neural TO

Rather than treating densities as elementwise design variables — the direct, mesh-coupled parameterization of standard TO — we can conceptualize the material distribution as a continuous field ρ(x)\boldsymbol{\rho}(\mathbf{x}) defined over the design domain x∈Ω\mathbf{x} \in\Omega, independent of the finite element discretization. This continuous perspective aligns naturally with the physics of the problem: material exists in space continuously, and the discrete finite element mesh is merely a computational tool required for the analysis. Neural fields provide an alternative parameterization, using neural networks to define smooth functions of spatial coordinates that can be evaluated at arbitrary locations [8]. In line with this approach, the density field can be expressed as

ρ(x,z;θ)=fθ(x,z(x))(6)\boldsymbol{\rho}(\mathbf{x}, \mathbf{z}; \boldsymbol{\theta}) = f_{\boldsymbol{\theta}}(\mathbf{x}, \mathbf{z}(\mathbf{x})) \tag*{(6)}

where fθf_{\boldsymbol{\theta}} is a neural network with learnable parameters θ\boldsymbol{\theta}, and z(x)\mathbf{z}(\mathbf{x}) is a conditioning input carrying problem-specific information. In this work, we reparameterized the density fields using sinusoidal representation networks (SIRENs) [16] with ResNet-like skip connections inspired by the Neural Implicit Flow framework [36]. SIREN uses periodic activation functions controlled by a single hyperparameter ω0\omega_0 to represent fine structural features, overcoming the spectral bias observed in standard multilayer perceptrons with ReLU activations [37]. Our architecture consisted of an input sine layer followed by two residual blocks of two sine layers each — five hidden layers of width 256 in total — and a linear output layer (as shown in Fig. 1), yielding a total of 264,449 learnable parameters. The network size was selected based on preliminary experiments to balance expressivity and computational cost. We found that the skip connections preserved the conditioning signal across network depth and stabilized the meta-training process.

Following Chen et al. [38], the conditioning input z(x)\mathbf{z}(\mathbf{x}) in our formulation corresponds to the normalized strain energy density of a reference design Π~(x)\widetilde{\boldsymbol{\Pi}}(\mathbf{x}). Here, the reference design is a uniform density distribution with values corresponding to the target volume fraction V∗V^*. Conditioning enables a single set of learned parameters θ\boldsymbol{\theta} to represent different structures depending on the problem specification. For each problem, we solved the equilibrium equation (2) for a reference design, and computed the elementwise strain energy densities as:

Πe=12ueTKeue,(7)\Pi_e = \frac{1}{2}\mathbf{u}_e^{\mathsf{T}}\mathbf{K}_e\mathbf{u}_e, \tag*{(7)}

where ue\mathbf{u}_e and Ke\mathbf{K}_e are the element displacement vector and stiffness matrix, respectively. These values were log-normalized to match the SIREN input range of [−1,1][-1,1]:

Π~e=2⋅log⁡(Πe)−min⁡e(log⁡(Πe))max⁡e(log⁡(Πe))−min⁡e(log⁡(Πe))−1(8)\widetilde{\Pi}_e = 2 \cdot\frac{\log(\Pi_e) - \min_e(\log(\Pi_e))}{\max_e(\log(\Pi_e)) - \min_e(\log(\Pi_e))} - 1 \tag*{(8)}

yielding the preprocessed conditioning values {Π~e}e=1N\{\widetilde{\Pi}_e\}_{e=1}^{N}.

To generate the discrete element densities required for the finite element solver, we sampled the network at NN element-specific inputs. These inputs were formed by concatenating the coordinates of element centroids xe\mathbf{x}_e with the corresponding preprocessed strain energy values Π~e\widetilde{\Pi}_e, yielding three-dimensional vectors [xe,Π~e]\left[\mathbf{x}_e, \widetilde{\Pi}_e\right].

The raw network outputs ρe\rho_e were amplified by a factor of 10 [39] and processed with two filters. First, we applied a linear density filter [40] with a radius of 132\frac{1}{32} of the domain size to ensure mesh-independent designs and prevent checkerboard artifacts. We denote the output of this transformation with ρˉe\bar{\rho}_e. Second, we applied the shifted sigmoid filter introduced by Hoyer et al. [32] to enforce the volume (3) and box (4) constraints:

Figure 1 (PDF p. 8). Residual SIREN architecture. The network maps element centroid coordinates xex_e and strain energy density Π~e\widetilde{\Pi}_e to element densities ρe\rho_e.

ρ~e=11+e−(ρˉe−b)\widetilde{\rho}_e = \frac{1}{1 + e^{-(\bar{\rho}_e-b)}}

where b∈Rb \in\mathbb{R} is a scalar bias computed at each iteration via bisection to satisfy the volume constraint exactly, i.e., the unique bb such that ∑e=1Nρ~eve=V∗\sum_{e=1}^{N}\widetilde{\rho}_e v_e = V^*. This transformation converts the original constrained optimization problem into an unconstrained one over the network parameters θ\boldsymbol{\theta}. While the original formulation (3) specifies an inequality constraint, this transformation enforces the volume constraint with equality. For compliance minimization, optimal designs use the full material budget, making these formulations equivalent in practice.

The complete neural topology optimization procedure is illustrated in Figure 2. At each iteration, the network produces a density field, which is filtered and evaluated with finite element analysis to compute the compliance c(θ)c(\boldsymbol{\theta}). The task loss is defined as the compliance normalized by that of a reference design under the same boundary conditions, L=c(θ)/cref\mathcal{L} = c(\boldsymbol{\theta})/c_{\mathrm{ref}}, enabling meaningful comparison across problems with different boundary conditions and volume fractions. Gradients of the loss with respect to the parameters of the neural field, ∇θL\nabla_{\theta}\mathcal{L}, are computed with automatic differentiation and used to update the network parameters via gradient descent. This optimization requires an initial set of parameters θ0\theta_0, whose choice can significantly influence both convergence speed and final design quality. Existing neural TO approaches rely on standard neural network initialization schemes [38] or pretrain the network to represent uniform density fields [11, 10]. These approaches are generic in that they do not account for the specific problem being solved.

Figure 2. Neural topology optimization (TO) procedure. At each iteration, the neural field produces a density field that is filtered and evaluated with finite element analysis to compute the compliance loss L\mathcal{L}.

Meta-neural TO

The topology optimization problems considered in this work share the same design domain, linear elastic material properties, and objective function, but differ in their boundary conditions and target volume fractions. We refer to each problem specification as a task Ti\mathcal{T}_i, and the set of tasks used for the learning process as the meta-training set. The procedure for pseudorandom tasks used in the experiments is described in § 5.4.

We propose to learn initial parameters θ0∗\theta_0^* that enable fast convergence across a set of related tasks through bilevel optimization. At the inner level, we consider a subset of tasks, and perform kk optimization steps to minimize their normalized compliance loss L\mathcal{L} (defined in § 5.1). We refer to this process as adaptation and denote the inner optimizer updates with UinnerTi\mathcal{U}_{\mathrm{inner}}^{\mathcal{T}_i}:

θi′←UinnerTi(θ0;α,k),(9)\theta_i' \leftarrow\mathcal{U}_{\mathrm{inner}}^{\mathcal{T}_i}(\theta_0; \alpha, k), \tag*{(9)}

where θ0\theta_0 is the initialization shared across tasks and α\alpha is the inner learning rate. At the outer level, we seek initial parameters θ0∗\theta_0^* that minimize the average post-adaptation loss across all MM meta-training tasks:

θ0∗=argmin⁡θ0 ∑i=1ML(θi′(θ0,Ti);Ti)⏟=:Lmeta(θ0).(10)\theta_0^* = \underset{\theta_0}{\operatorname{argmin}}\ \underbrace{\sum_{i=1}^{M} \mathcal{L}(\theta_i'(\theta_0,\mathcal{T}_i);\mathcal{T}_i)}_{=:\mathcal{L}_{\mathrm{meta}}(\theta_0)}. \tag*{(10)}

We refer to the outer objective as the meta-loss and denote it Lmeta(θ0)\mathcal{L}_{\mathrm{meta}}(\theta_0), dropping the explicit dependence on the tasks for notational simplicity. We explored two gradient-based meta-learning algorithms for solving the outer optimization in (10): MAML [14] and Reptile [15]. MAML computes the meta-gradients ∇θ0Lmeta\nabla_{\theta_0}\mathcal{L}_{\mathrm{meta}} exactly. Since the meta-loss is an implicit function of the inner-loop gradients, this calculation relied on second-order gradients, which were prohibitively expensive to compute and store in our context. Consequently, we adopted Reptile for all experiments in this work. Reptile approximates the meta-gradients as the difference between the current initialization and the adapted parameters, averaged over a batch of BB tasks:

∇θ0Lmeta≈1B∑i=1B(θ0−θi′)(11)\nabla_{\theta_0}\mathcal{L}_{\mathrm{meta}} \approx\frac{1}{B}\sum_{i=1}^{B}(\theta_0-\theta_i') \tag*{(11)}

and updates the shared initialization by applying a single step of an outer optimizer Uouter\mathcal{U}_{\mathrm{outer}}:

θ0←Uouter(θ0;η)(12)\boldsymbol{\theta}_0 \leftarrow\mathcal{U}_{\mathrm{outer}}(\boldsymbol{\theta}_0;\eta) \tag*{(12)}

where η\eta is the outer learning rate.

Unlike supervised approaches that require precomputed datasets of optimized structures, our method accumulates design experience through the optimization process itself. The outer loop evaluates initialization quality based on post-adaptation performance, learning from partial optimizations rather than from final designs. The meta-learning procedure is illustrated in Figure 3 with the corresponding pseudocode in Algorithm 1.

Figure 3. Meta-neural topology optimization (TO) framework. The outer loop learns a shared initialization θ0∗\boldsymbol{\theta}_0^* from partial optimization trajectories across tasks, enabling fast adaptation to new design problems.

Each inner-loop iteration in Algorithm 1 (line 8) maps to one cycle of the neural TO procedure shown in Figure 2: the neural field produces a density field, which is filtered and evaluated via finite element analysis to compute the compliance loss; the network parameters are then updated based on the gradients. After kk such iterations, the adapted parameters θi′\boldsymbol{\theta}'_i are compared to the shared initialization θ0\boldsymbol{\theta}_0 to form the Reptile gradient estimate (11), which drives the outer-loop update of θ0\boldsymbol{\theta}_0.

We chose the Adam optimizer [41] for both the inner and outer loop updates. In line with previous works in neural TO [9, 10], we observed that simple gradient descent was unable to navigate nonconvex loss landscapes induced by reparameterization. The inner and outer learning rates, α\alpha and η\eta, and the frequency parameter ω0\omega_0 of the Residual SIREN model were selected through a grid search on a held-out validation set, as detailed in Section A, with the final selection summarized in Table 1. The inner loop consisted of k=10k = 10 iterations, striking a balance between computational efficiency and effective task-specific adaptation. We trained Reptile for 6000 outer iterations with a meta-batch size of B=5B = 5, chosen so

ParameterSymbolValue
Meta-neural TO
Meta-learning algorithm—Reptile
Outer optimizer—Adam
Outer learning rateη\eta10−610^{-6}
Outer iterations—6000
Meta-batch sizeBB5
Inner optimizer—Adam
Inner learning rateα\alpha10−410^{-4}
Inner stepskk10
Frequency parameterω0\omega_060
Neural TO
Frequency parameterω0\omega_030
Learning rate—10−410^{-4}

Table 1. Hyperparameters for meta-neural TO and the neural TO baseline. Values were selected based on validation performance across all tested configurations.

Algorithm 1 Meta-neural topology optimization with Reptile

Require: Meta-training set Str={Ti}i=1M\mathcal{S}^{\mathrm{tr}} = \{\mathcal{T}_i\}_{i=1}^{M}

Require: Inner learning rate α\alpha, outer learning rate η\eta, inner steps kk, meta-batch size B\mathcal{B}

1: Initialize parameters θ0\boldsymbol{\theta}_0

2: Initialize outer optimizer (Adam with learning rate η\eta)

3: for number of outer steps do

4:     Sample a batch of tasks B={Ti}i=1B∼Str\mathcal{B} = \{\mathcal{T}_i\}_{i=1}^{B} \sim\mathcal{S}^{\mathrm{tr}}

5:     Initialize meta-gradient accumulator g←0\boldsymbol{g} \leftarrow\boldsymbol{0}

6:     for each task Ti∈B\mathcal{T}_i \in\mathcal{B} do

7:         Initialize inner optimizer (Adam with learning rate α\alpha)

8:         Adapt to task: θi′←Uinner(θ0;α,k)\boldsymbol{\theta}'_i \leftarrow\mathcal{U}_{\mathrm{inner}}(\boldsymbol{\theta}_0; \alpha, k) // (9)

9:         Accumulate: g←g+(θ0−θi′)\boldsymbol{g} \leftarrow\boldsymbol{g} + (\boldsymbol{\theta}_0 - \boldsymbol{\theta}'_i)

10:     end for

11:     Compute meta-gradient: ∇θ0Lmeta←1B⋅g\nabla_{\boldsymbol{\theta}_0}\mathcal{L}_{\mathrm{meta}} \leftarrow\frac{1}{B} \cdot\boldsymbol{g} // (11)

12:     Update initialization: θ0←Uouter(θ0;η)\boldsymbol{\theta}_0 \leftarrow\mathcal{U}_{\mathrm{outer}}(\boldsymbol{\theta}_0; \eta) // (12)

13: end for

14: return Learned initialization θ0∗\boldsymbol{\theta}_0^*

that each task in the meta-training dataset was sampled exactly once per epoch. Extending training beyond this point led to meta-overfitting, where the initialization specialized on a subset of tasks at the cost of generalization [13]. The evolution of meta losses across the meta-training for the selected hyperparameters is shown in Figure 4.

Figure 4. Meta-training loss curve. Evolution of the meta-loss for the selected hyperparameters (ω0=60.0\omega_0 = 60.0, η=10−6\eta= 10^{-6}, α=10−4\alpha= 10^{-4}).

At meta-testing, we optimized θ\boldsymbol{\theta} for each task for a minimum of 10 and a maximum of 200 iterations, starting from the learned initialization θ0∗\boldsymbol{\theta}_0^*. We applied a stopping criterion based on the change in the loss between consecutive iterations:

∣L(θj)−L(θj−1)∣<ϵ(1+∣L(θj−1)∣)\lvert\mathcal{L}(\theta_j) - \mathcal{L}(\theta_{j-1}) \rvert< \epsilon(1 + \lvert\mathcal{L}(\theta_{j-1}) \rvert)

where L(θj)\mathcal{L}(\theta_j) represents the loss at optimization iteration jj and ϵ=10−5\epsilon= 10^{-5}. This criterion accounts for the scale of the objective function and is well-suited for stochastic optimizers, such as Adam [42]. Throughout this paper, we use “convergence” to refer to reaching the stopping criterion, as opposed to the strict fulfillment of the optimality criterion. To ensure fair comparisons, we apply volume-preserving thresholding [4, 43] to the final designs. As explained in Section B, thresholding had a disproportionately positive effect on standard TO designs compared to neural TO designs.

Task generation

To enable meta-learning, we constructed datasets of pseudorandom TO tasks, each defined by a set of boundary conditions and a target volume fraction. All tasks were defined on a square design domain discretized by a structured grid of 64×6464 \times64 elements. We generated 3000 meta-training tasks with point loads and supports using the method proposed in Sosnovik and Oseledets [27]. Each task was constructed by randomly sampling the numbers and positions of boundary and load conditions. The number of support locations and loads per task were sampled from Poisson distributions (mean of μ=1\mu= 1 for supports, μ=2\mu= 2 for loads), with minimum counts enforced to ensure well-posed problems (at least two support locations and one load per task). Supports were placed preferentially along the domain edges. To enforce a minimum separation between boundary conditions, the sampling probability was modulated based on the distance from existing supports. Target volume fractions were sampled uniformly from [0.1,0.5][0.1, 0.5], and load directions uniformly from [0,2π][0, 2\pi] radians.

We constructed three test datasets of 1000 tasks each to evaluate different levels of generalization: (i) an in-distribution dataset with point loads and supports matching the training distribution, (ii) an out-of-distribution dataset where a subset of boundary conditions were distributed along randomly generated line segments on the domain boundary, following Mazé and Ahmed [44], and (iii) a cross-resolution dataset with point loads and supports on a 256×256256 \times256 discretization. Further, we generated 100 tasks for hyperparameter selection using the same generation process as for the meta-training data. Examples of boundary conditions in in-distribution and out-of-distribution datasets are shown in Figure 5. When generating the datasets, we performed one finite element analysis per task to calculate the strain energy field Π~\tilde{\Pi} and the reference compliance crefc_{\mathrm{ref}}. Conveniently, this enabled us to filter out tasks with rigid-body modes by identifying singular matrix errors in the equilibrium solves.

Figure 5. Example boundary conditions. Tasks from the (a) in-distribution dataset with point loads and supports, and (b) out-of-distribution dataset with distributed boundary conditions along line segments.

Results

We evaluated our approach on three test datasets, examining (i) whether meta-learned initializations can consistently accelerate convergence, (ii) how they affect the final design compliance, and (iii) how they generalize to tasks outside of the training distribution. Two additional datasets evaluating robustness to domain geometry and compound distribution shifts are reported in Section F. We evaluated meta-neural TO against two baselines: (i) standard density-based TO without neural reparameterization solved using the globally-convergent version of the MMA algorithm (GCMMA) with default NLopt settings [45] (no hyperparameter tuning), and (ii) neural TO (as described in § 5.2) with a principled SIREN initialization scheme [16] and hyperparameters tuned on the validation set (ω0=30\omega_0 = 30, the best-performing configuration for standard neural TO; see Section A). To ensure fair comparison, we report single runs for meta-neural TO and standard TO, as their performance is deterministic given the strain energy conditioning and problem setup, respectively. To present the strongest possible baseline, we ran neural TO with three different random seeds and reported the best-case results, i.e., those from the seed that required the fewest iterations to converge on the in-distribution dataset. The baselines offer complementary perspectives for assessing our method. Neural TO helps to isolate the effect of meta-learned initialization, while standard TO provides context for the practical relevance of our approach. We compared the methods using performance profiles [46], as shown in Figure 6. For each task, we identified the best-performing method(s) and assessed the performance of other methods relative to it. The profiles illustrate the fraction of tasks each method solves within a tolerance of the best result. For example, a performance ratio of 0.9 at the tolerance of 1.5 indicates that a given method solves 90% of the test tasks within a 50% error of the best performer for each task. The profiles reveal how often an algorithm is the top performer (tolerance = 1) and how closely it trails the leader (at tolerance > 1). We direct the readers to Dolan and Moré [46] for methodological details on performance profiles and Rojas-Labanda and Stolpe [47] for an application in TO.

Figure 6. Performance profiles for meta-neural TO and baselines. Top row (a–c): number of iterations to convergence. Bottom row (d–f): thresholded design compliance. Columns correspond to in-distribution, out-of-distribution, and cross-resolution experiments, respectively.

In-distribution

Meta-neural TO converged in the fewest iterations in 57.6% of in-distribution tasks, followed by standard TO at 42.3% and neural TO at 11.7% (Figure 6a). The sum of performance ratios exceeds 100% here because of ties, primarily on tasks where all methods use the maximum budget of 200 iterations and therefore tie for the fastest convergence. This only occurred for 48 tasks, confirming that the 200-iteration budget was sufficient for performance comparisons. On average, meta-neural TO required 103.65 iterations, standard TO 111.26, and neural TO 149.56 (see Table 2). These results indicate that meta-neural TO not only matches the optimization efficiency of the conventional approach but can surpass it.

In-distributionOut-of-distributionCross-resolution
Standard TO111.26110.89130.03
Neural TO149.56154.67152.04
Meta-neural TO103.65126.57100.90

Table 2. Average number of iterations to convergence across test datasets. Bold values indicate the lowest average for each scenario.

Figure 7 shows how the average loss over the in-distribution tasks evolves across meta-testing iterations. While both neural approaches demonstrate rapid initial loss reduction, meta-neural TO starts at a lower loss value and maintains the lowest loss throughout the optimization. Standard TO requires a few iterations before significant loss reduction occurs. The initial gradual convergence is attributed to the default parameters in GCMMA, which prioritize robustness over aggressive early optimization steps. Eventually, all three methods reach low average loss values, confirming that each approach can reliably find high-quality solutions across the task distribution.

Average convergence curves

Figure 7. Average convergence curves. Mean normalized compliance loss over the in-distribution test tasks as a function of optimization iteration.

While standard TO found the best thresholded designs in 63.1% of tasks, compared to 20.3% for meta-neural TO and 16.6% for neural TO (Figure 6d), all methods performed comparably within a 5% tolerance, with performance ratios of 83.9%, 80.9%, and 80.4% for standard TO, meta-neural TO and neural TO, respectively. Meta-neural TO frequently produced designs with quality comparable to standard TO. In fact, the analysis in Section B shows that meta-neural TO found the best continuous designs in 69.1% of tasks.

Out-of-distribution

Meta-neural TO was fastest on 39.3% of out-of-distribution tasks, outperforming neural TO at 12.5% but trailing standard TO at 59.9% (Figure 6b). Meta-neural TO required an average of 126.57 iterations — fewer than neural TO (154.67) but more than standard TO (110.89) (see Table 2). The quality of thresholded designs showed a similar trend (Figure 6e) — meta-neural TO found superior designs in 24.6% of tasks versus standard TO’s 47.6%, with both methods achieving similar success rates at a 5% tolerance.

Cross-resolution

Meta-neural TO excels in cross-resolution experiments, where it was applied to tasks with finer discretization than used at meta-training. It required the fewest iterations in 74.1% of tasks compared to 25.7% for standard TO and 10.8% for neural TO (Figure 6c). On average, meta-neural TO converged in 100.90 iterations, standard TO in 130.03, and neural TO in 152.04 (see Table 2) — the learned initialization reduced the average number of iterations by 33.6% compared to neural TO. Figure 8 showcases examples from our cross-resolution test set, where meta-neural TO consistently achieves the lowest compliance values while requiring significantly fewer iterations than both baseline methods.

Figure 8. Cross-resolution design examples. Each row shows a test case with optimized structures and performance metrics reported as (normalized compliance before/after thresholding, iterations to convergence).

This experiment highlights the advantageous scaling of neural representations — while standard TO must handle 16 times more design variables, neural TO benefits from a mesh-independent parameterization. The improved efficiency did not come at the cost of final design quality, as meta-neural TO and standard TO found competitive final thresholded designs, winning in 41.0% and 48.5% of tasks, respectively (Figure 6f). At a 5% tolerance, meta-neural TO achieved a performance ratio of 87.7%, surpassing standard TO’s 83.6%. These results demonstrate an effective generalization of meta-learned initializations to discretizations finer than those seen during meta-training.

Representative designs

Figure 9 illustrates the convergence curves and final designs (before thresholding) for a set of representative in-distribution tasks. These tasks correspond to the best, worst, and typical (median) improvements in the final design quality of meta-neural TO compared to neural TO. The best and typical (median) designs achieved 37.72% and 1.50% improvements in compliance, respectively. Notably, in both cases, the meta-learned initialization corresponds to high-quality initial designs (as indicated by low compliance at the 0th iteration). The final designs from meta-neural TO exhibited finer structural features, most likely due to the higher frequency parameter used in the model (ω0=60\omega_0 = 60 for meta-neural TO versus ω0=30\omega_0 = 30 for neural TO). The worst design showed a drastic 149.8% increase in compliance. In this case, the meta-learned initial design underperformed a generic initialization and biased the optimization towards a poor local minimum. Notably, neither of the two approaches converged within the 200-iteration budget. As shown in Figure 10, while meta-learned initializations generally improved performance across volume fractions (positive median values), tasks with lower target volumes (0.1-0.2) exhibited higher variability, with both the largest potential improvements and deterioration in final design quality.

Figure 9. Representative in-distribution tasks. Convergence curves and final designs for tasks with the best, worst, and median improvement in continuous compliance from meta-neural TO over neural TO.

Scatter plot of compliance improvement by target volume fraction

Figure 10. Compliance improvement by volume fraction. Improvement of meta-neural TO over neural TO on in-distribution tasks, grouped by target volume fraction. Black lines indicate bin medians.

Strain energy as a design prior

The analysis of initial designs from meta-neural TO across different tasks reveals an interesting connection. As shown in Figure 11, the initial density fields closely mirror the ones obtained by directly applying the volume and density filters to the input strain energy fields Π~(x)\widetilde{\Pi}(\mathbf{x}). While the network is conditioned on strain energy fields, it was not explicitly trained to reproduce them as initial designs. Instead, the meta-learning process discovered that directly mapping the filtered strain energy patterns provided an effective initialization strategy across diverse optimization tasks.

Figure 11. Meta-learned initial designs as strain energy priors. Initial designs from meta-neural TO (bottom) closely mirror the filtered strain energy density fields of uniform density reference designs (top).

This discovery has a clear physical interpretation. The sensitivity of the compliance objective with respect to element densities is proportional to the local strain energy density:

∂c∂ρe=−pρep−1ueKe(ρe)ue(13)\frac{\partial c}{\partial\rho_e} = -p \rho_e^{p-1} \mathbf{u}_e \mathbf{K}_e(\rho_e) \mathbf{u}_e \tag*{(13)}

Consequently, placing material where strain energy density is high and removing it where strain energy density is low is a natural strategy for compliance minimization. This principle underlies several established TO approaches, including the bubble method [48] and bi-directional evolutionary structural optimization [49]. One might then ask why standard TO does not simply initialize from strain energy patterns. The reason is that MMA makes conservative local updates that cannot easily escape an early material layout once the design approaches a solid/void state [4]. This limitation does not apply to neural TO, where the continuous parameterization permits larger structural rearrangements, a property that the meta-learner exploits. The role of strain energy conditioning is examined through an ablation in Section D.

Given that meta-learning converges to strain-energy-based initializations, a natural question is whether this prior can be encoded more directly. To verify this finding, we performed an additional experiment (see Section C), where we pretrained a neural field to perform an identity mapping on the strain energy input across training tasks, i.e., fθ(x,Π~(x))≈Π~(x)f_{\theta}(\mathbf{x}, \tilde{\Pi}(\mathbf{x})) \approx\tilde{\Pi}(\mathbf{x}). Surprisingly, this simple approach outperformed meta-neural TO in both convergence speed and final design quality. The improvement can be attributed to the directness of the pretraining objective: while meta-learning must implicitly discover the relevance of strain energy patterns through thousands of partial optimizations, pretraining encodes this mapping exactly, producing initial designs that more faithfully reproduce the filtered strain energy fields. In effect, pretraining provides the same physical prior but with less noise from the meta-optimization process.

Computational complexity

Meta-neural TO and neural TO have an identical per-iteration cost at test time. Both evaluate the same neural field architecture, followed by finite element analysis, filtering, and Adam optimizer updates. Consequently, the iteration counts directly reflect the differences in total computation time. While standard TO does not require an evaluation of a neural field, it uses a more complex GCMMA optimizer. We evaluated the wall-clock times across 200 iterations for three classical benchmark cases: MBB, cantilever, and tensile beams [47, 4], and collected the average per-iteration costs in Table 3.

Standard TONeural TOMeta-neural TO
Time per iteration (64×6464 \times64) [ms]46.1162.49same as neural TO
Time per iteration (256×256256 \times256) [ms]723.13672.90same as neural TO
Avg. iterations (in-dist., 64×6464 \times64)111.26149.56103.65
Meta-training time——3h 35min

Table 3. Computational complexity comparison. Wall-clock times were measured on a single NVIDIA A100 GPU. Per-iteration times are averaged over 200 iterations across three benchmark tasks.

The per-iteration cost of neural TO is approximately 1.36×1.36\times that of standard TO on the 64×6464 \times64 mesh. However, on the 256×256256 \times256 mesh, neural TO becomes faster (0.93×0.93 \times the cost of standard TO). This performance shift occurs because the computational cost of the GCMMA optimizer update scales less favorably than the neural network evaluation at higher resolutions. While our implementations are not fully optimized, these timings serve as a valid relative comparison.

The meta-training represents a one-time investment of 3 hours and 35 minutes on a single NVIDIA A100 GPU. The average throughput during meta-training (≈43\approx43 ms/iter) is faster than the instrumented test-time measurement (≈62\approx62 ms/iter) due to the use of just-in-time compiled loops and reduced logging overhead. This cost is amortized over all subsequent test problems. Notably, this cost is equivalent to approximately 1500 full optimizations on the 64×6464 \times64 mesh, which is significantly lower than the cost of preparing datasets for supervised deep learning approaches. For context, comparable data-driven methods typically rely on generating datasets ranging from 30,000 to over 2,200,000 pre-optimized samples [26].

Discussion

We introduced meta-neural TO, a framework that accumulates and transfers design experience across topology optimization problems via meta-learning. Unlike pretraining approaches [9, 11] and direct-design methods [30], our method learns from partial optimizations rather than precomputed datasets, making it readily extensible to different optimizers and parameterizations. The cross-resolution experiments further show that this knowledge transfers across mesh discretizations, a property unique to neural parameterizations. The two additional datasets of Section F delimit how far this transfer reaches: a change in domain aspect ratio leaves the iteration advantage intact (fewest iterations on 63.6% of tasks), whereas combining distributed boundary conditions with a finer mesh erodes it (40.9% versus 56.0% for standard TO) — yet even in that hardest setting meta-neural TO still yields the best thresholded designs.

A central finding of this work is that strain energy conditioning is essential for effective knowledge transfer. The meta-learning process independently converged to strain-energy-based initializations without being explicitly guided to do so, and the theoretical justification via the compliance sensitivity (13) confirms why this strategy is effective. That the same initializations also accelerate standard TO (Section E) suggests the discovery of optimizer-agnostic structural priors with potential impact beyond the neural TO setting.

The method has some practical limitations. Meta-training demands substantial computation. tional resources, though training on coarse meshes mitigates this cost. Additionally, generating a diverse task distribution requires careful consideration. For compliance minimization, pseudorandom sampling of boundary conditions suffices, but for more complex physics, the challenge grows considerably. In fracture mechanics, for instance, one must ensure that generated tasks actually exhibit crack initiation and propagation rather than trivial elastic responses—curating such a dataset is itself a nontrivial problem. Furthermore, the meta-learned initialization can bias the optimizer towards poor local minima, as shown by the worst-case in-distribution task, and performance degrades when test problems deviate substantially from the training distribution.

The practical value of faster convergence is greatest in domains where families of related problems are solved repeatedly. In aerospace and automotive engineering, components are routinely optimized under varying boundary conditions but with similar underlying physics. The topology optimization of wheel designs [50], where load cases and symmetry constraints vary across otherwise similar structures, is one such example. The method is also well-suited for rapid prototyping and high-resolution problems, where per-iteration cost is high, and iteration savings translate directly into wall-clock gains. Promising directions include per-parameter learning rate adaptation via Meta-SGD [51], boundary condition similarity metrics to guide task selection [52], multi-fidelity strategies building on cross-resolution transfer, and extensions to more complex physics such as plasticity, fracture, or thermal problems—where initialization is known to strongly influence which local optimum the solver reaches [53].

CRediT authorship contribution statement

Igor Kuszczak: Conceptualization, Methodology, Investigation, Software, Writing – original draft. Gaweł Kuś: Software, Data curation, Writing – review & editing. Federico Bosi: Supervision, Writing – review & editing, Funding acquisition. Miguel A. Bessa: Conceptualization, Supervision, Writing – review & editing, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This research was supported by the European Union’s Horizon 2020 Research and Innovation programme through the Marie Skłodowska-Curie grant (Agreement No. 956547-LIGHTEN). M.A.B. would like to acknowledge that this effort was undertaken in part with the support from the Department of the Navy, Office of Naval Research, award number N00014-23-1-2688. G.K. would like to acknowledge that this effort was undertaken in part with the support from the Department of the Navy, Office of Naval Research, award number N00014-21-1-2670. I.K. acknowledges Suryanarayanan Manoj Sanu and Leon Herrmann for their valuable feedback on the manuscript. Special thanks to Louis Kirsch for the discussions that laid the foundation for this work.

Data availability

The code used to generate the results presented in this paper is available at https://github.com/bessagroup/metatopia. The complete dataset used in this study is publicly accessible and can be downloaded from Zenodo (DOI: 10.5281/zenodo.15172934).

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the authors used Claude (Anthropic) to assist with proofreading and language editing, including grammar correction, clarity improvements, and consistency checks across the manuscript. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Hyperparameter selection

Building on previous studies on meta-learning for neural fields [17, 16], we conducted a focused grid search over the key model parameters — inner and outer learning rates, and frequency parameter ω0\omega_0. We performed full meta-training for each hyperparameter combination and evaluated the average normalized compliance loss after 10 iterations on a held-out validation set (Table A.1).

Modelω0\omega_0η\etaα\alphaValidation loss
MetaStandard
Residual SIREN30.010−510^{-5}10−410^{-4}0.1440.114
Residual SIREN30.010−610^{-6}10−410^{-4}0.1030.114
Residual SIREN30.010−510^{-5}10−510^{-5}0.1430.418
Residual SIREN30.010−610^{-6}10−510^{-5}0.5130.418
Residual SIREN60.010−510^{-5}10−410^{-4}0.1460.128
Residual SIREN60.010−610^{-6}10−410^{-4}0.0870.128
Residual SIREN60.010−510^{-5}10−510^{-5}0.2830.487
Residual SIREN60.010−610^{-6}10−510^{-5}0.1250.487

Table A.1. Validation loss comparison across different hyperparameter configurations. Configurations where meta-neural TO outperforms standard neural TO are highlighted in green. Bold values represent the best-performing configurations.

As summarized in Table A.1, meta-neural TO outperformed neural TO with standard initialization in 5 out of 8 presented configurations, with the lowest validation loss at ω0=60.0\omega_0 = 60.0, outer learning rate of 10−610^{-6}, and inner learning rate of 10−410^{-4}. Notably, standard neural TO performed best with ω0=30.0\omega_0 = 30.0. This finding aligns with previous work on coordinate-based neural representations [17, 54], which shows that meta-learned models can benefit from higher ω0\omega_0 values, suggesting that learned initialization enables an effective use of higher-frequency components.

Effect of thresholding on design quality

The performance profiles in Fig. B.1 reveal an interesting pattern where meta-neural TO finds superior continuous designs for the majority of tasks across all test datasets. This is most pronounced in the cross-resolution experiment, where meta-neural TO dominates on 81.0% of tasks. Note that these results complement the performance profiles in Figure 6. We observed that thresholding had a disproportionately positive effect on standard TO designs, reducing this apparent advantage. Here, we report the average improvements in compliance from thresholding for meta-neural TO and standard TO designs. For the calculation, we exclude the tasks where thresholding led to a drastic (>50%> 50\%) quality degradation. Thresholding improved the compliance of standard TO designs by 6.63% (in-distribution) and 5.68% (out-of-distribution). For meta-neural TO designs, thresholding improved compliance by 1.97% (in-distribution) and 1.74% (out-of-distribution). We found that neural TO approaches generally produced designs with fewer intermediate-density elements, leading to smaller gains from thresholding. We attribute this effect to the sigmoidal volume-preserving filter.

Figure B.1. Continuous design compliance. Performance profiles comparing continuous (not thresholded) design compliance for (a) in-distribution, (b) out-of-distribution, and (c) cross-resolution experiments.

Alternative pretraining strategy

To better understand the relationship between strain energy fields and effective initial designs, we compared our meta-learning approach against a simple pretraining strategy. We trained the neural field to directly output strain energy density fields (one of its inputs) using mean squared error loss over 100 epochs (understood as sweeps through the meta-training dataset). Performance profiles in Fig. C.1 show that this approach offers moderate efficiency improvements over the meta-learned initialization, requiring the fewest iterations in 40.7% of tasks compared to meta-neural TO’s 37.4%. Further, the approach improves continuous and thresholded compliance across tasks, narrowing the quality gap between neural and standard TO.

This approach’s strong performance validates a key insight of our meta-learning framework: the critical importance of strain energy patterns in generating effective initial designs. Meta-learning naturally discovered this relationship without explicit supervision, demonstrating its ability to uncover meaningful physical principles from the optimization process alone. This discovery mechanism makes our meta-learning approach particularly promising for applications where the relevant physical patterns are less well understood or where multiple fields need to be considered simultaneously.

Figure C.1. Pretraining versus meta-learning. Performance profiles comparing iterations to convergence (a), continuous compliance (b), and thresholded compliance (c) for the pretrained neural TO, meta-neural TO, and standard TO on in-distribution tasks.

Meta-neural TO without conditioning

The proposed meta-learning method relies on strain energy fields to generate task-specific initial designs. Here, we examine how removing this conditioning mechanism affects the meta-learning process and resulting performance. Without access to boundary condition information, the meta-learning objective fundamentally shifts. Instead of learning to map a physical field into promising initial designs, the model must discover a universal initial design that can rapidly adapt to diverse topologies while learning about problem physics solely through the loss function.

Performance profiles in Fig. D.1 reveal that removing strain energy conditioning impacts both the convergence speed and the final design quality. While meta-learned initialization still outperforms neural TO with random initialization (both without conditioning), the improvements are substantially smaller than in the conditioned case. Notably, the unconditioned meta-learned strategy is less efficient than standard TO for in-distribution tasks. Analysis of initial designs in Fig. D.2 shows that the unconditioned model learns to concentrate material along domain boundaries, mirroring the probability distribution used for sampling boundary conditions in the meta-training set. This suggests that even without explicit physical information, meta-learning can discover useful structural priors from the statistical patterns in the training data.

Figure D.1. Effect of removing strain energy conditioning. Performance profiles comparing iterations to convergence (a), continuous compliance (b), and thresholded compliance (c) for standard TO, neural TO, and meta-neural TO, with both neural methods evaluated without strain energy conditioning.

Figure D.2. Unconditioned meta-learned initial designs. Without strain energy conditioning, the model concentrates material along domain boundaries, reflecting the statistical distribution of boundary conditions in the training set.

Can learned initializations benefit standard TO?

Finally, we investigated whether the initial designs from meta-neural TO can accelerate standard TO. The neural TO designs satisfy the volume constraint by construction and thus can be used as initial guesses for the MMA optimizer. The experiments with initial designs from conditioned and unconditioned meta-neural TO reveal an interesting dichotomy. As illustrated in Fig. E.1, initializing MMA with designs from the strain-energy-conditioned meta-neural TO improves optimization efficiency. In contrast, initial designs from the unconditioned model lead to worse performance.

This distinction highlights two key aspects of meta-learning in TO. First, the meta-learning process inherently considers optimizer dynamics when searching for improved initializations, making direct transfer between different optimizers challenging. This explains why unconditioned initializations, while effective for neural TO, do not benefit standard TO. Second, when conditioned on strain energy density, the meta-learned model discovers physically meaningful priors for material placement that prove beneficial regardless of the optimization method. These findings underscore both the optimizer-specific nature of meta-learned initialization strategies and the potential for physics-informed conditioning to enable cross-optimizer knowledge transfer.

Figure E.1. Cross-optimizer transfer. Performance profiles comparing standard TO initialized from uniform, conditioned, and unconditioned meta-learned designs in terms of iterations to convergence (a), continuous compliance (b), and thresholded compliance (c).

Evaluation on additional datasets

We introduced two additional datasets to assess how well our method generalizes under domain shift: a cross-aspect dataset featuring a rectangular design domain, and a compound-shift dataset that combines out-of-distribution boundary conditions with higher resolution discretizations from the cross-resolution dataset. Together with the experiments in the main text, this extended the evaluation to five datasets spanning a range of aspect ratios, boundary condition types, and mesh resolutions. All experiments follow the meta-testing protocol outlined in the main text. Each method runs for a minimum of 10 and a maximum of 200 iterations subject to the relative-change stopping criterion (ϵ=10−5\epsilon= 10^{-5}), and all three methods are compared using performance profiles on the same set of tasks.

Cross-aspect

The cross-aspect dataset evaluates robustness to changes in domain geometry. All main-text experiments used a square design domain; here, we constructed 1000 tasks on a rectangular 128×64128 \times64 domain with point loads and supports, following the same generation procedure as the in-distribution dataset (§ 5.4). Although the neural-field parameterization naturally accommodates different domain shapes through its continuous spatial coordinates, it is not guaranteed that the learned initialization remains effective under this geometric shift.

Figure F.1. Performance profiles on the cross-aspect dataset. Iterations to convergence (a), continuous compliance (b), and thresholded compliance (c).

Performance profiles for the cross-aspect dataset are shown in Fig. F.1. Meta-neural TO required the fewest iterations in 63.6% of tasks, ahead of standard TO at 31.5% and neural TO at 11.6% (Fig. F.1a), and produced the best continuous designs in 59.1% of tasks, ahead of standard TO (26.2%) and neural TO (14.7%) (Fig. F.1b). Although the ordering matches the in-distribution experiment, two differences stand out. First, meta-neural TO’s iteration lead grows relative to in-distribution (from 57.6% to 63.6%) while standard TO’s share drops (from 42.3% to 31.5%). As in the cross-resolution experiment, standard TO must now optimize over twice the number of design variables, whereas the neural parameterization is mesh-independent.

Second, the thresholded-compliance gap widens in favor of standard TO (78.5%, up from 63.1% in-distribution), with meta-neural TO and neural TO dropping to 13.8% and 7.7% (from 20.3% and 16.6%, respectively, Fig. F.1c). Because both neural methods move in the same direction, we attribute this shift to the neural parameterization itself rather than to meta-learning. As discussed in Section B, neural-field designs contain fewer intermediate-density elements than density-based ones and therefore benefit less from the volume-preserving filter — an effect that widens under the geometric shift.

Compound shift

The compound-shift dataset consists of 1000 tasks with distributed boundary conditions (as in the out-of-distribution dataset) on a 256×256256 \times256 mesh (as in the cross-resolution dataset), exercising both distribution shifts simultaneously.

Figure F.2. Performance profiles on the compound-shift dataset. Iterations to convergence (a), continuous compliance (b), and thresholded compliance (c).

Performance profiles for the compound-shift dataset are shown in Fig. F.2. Standard TO required the fewest iterations in 56.0% of tasks, followed by meta-neural TO at 40.9% and neural TO at 8.6% (Fig. F.2a). Meta-neural TO produced the best continuous designs in 65.4% of tasks, ahead of neural TO at 20.1% and standard TO at 14.5% (Fig. F.2b), and the best thresholded designs in 42.7% of tasks, ahead of standard TO at 32.7% and neural TO at 24.6% (Fig. F.2c).

The iteration profile resembles the out-of-distribution experiment (meta-neural TO 39.3%, standard TO 59.9%) far more than the cross-resolution experiment (meta-neural TO 74.1%).

The distributed-boundary-condition shift dominates iteration behavior, and adding the resolution shift on top does not recover the speed advantage that meta-neural TO enjoys on cross-resolution alone.

The thresholded compliance profile is more striking. In both the out-of-distribution and cross-resolution experiments of the main text, standard TO won the majority of thresholded comparisons at tolerance 1 (47.6% and 48.5%, respectively), consistent with the broader pattern that standard TO benefits more from thresholding (Section B). On the compound-shift dataset, this ordering reverses, with meta-neural TO winning on 42.7% of tasks, standard TO 32.7%, and neural TO 24.6%. Interestingly, on 9.2% of tasks, thresholded standard TO designs perform at least ten times worse than the best-in-task design, compared to 5.8% for neural TO and 3.1% for meta-neural TO. This severe degradation is linked to a recurring failure mode where standard TO produces thin, elongated members with diffuse boundaries that physically disconnect upon thresholding, resulting in degenerate structures. Meta-neural TO’s lead over standard TO is preserved at a 5% tolerance (meta-neural TO 74.9%, neural TO 69.3%, standard TO 60.2%).

References

  1. [1]M. Albrechtsen, B. Vosoughi Lahijani, R. E. Christiansen, V. T. H. Nguyen, L. N. Casses, S. E. Hansen, N. Stenger, O. Sigmund, H. Jansen, J. Mørk, S. Stobbe, Nanometer-scale photon confinement in topology-optimized dielectric cavities, Nature Communications 13 (2022) 6281. doi:10.1038/s41467-022-33874-w.DOI
  2. [2]N. Aage, E. Andreassen, B. S. Lazarov, O. Sigmund, Giga-voxel computational morphogenesis for structural design, Nature 550 (2017) 84–86. doi:10.1038/nature23911.DOI
  3. [3]M. P. Bendsøe, N. Kikuchi, Generating optimal topologies in structural design using a homogenization method, Computer Methods in Applied Mechanics and Engineering 71 (1988) 197–224. doi:10.1016/0045-7825(88)90086-2.DOI
  4. [4]O. Sigmund, K. Maute, Topology optimization approaches, Structural and Multidisciplinary Optimization 48 (2013) 1031–1055. doi:10.1007/s00158-013-0978-6.DOI
  5. [5]M. P. Bendsøe, Optimal shape design as a material distribution problem, Structural optimization 1 (1989) 193–202. doi:10.1007/BF01650949.DOI
  6. [6]K. Svanberg, The method of moving asymptotes—a new method for structural optimization, International Journal for Numerical Methods in Engineering 24 (1987) 359–373. doi:10.1002/nme.1620240207.DOI
  7. [7]D. A. White, Y. Choi, J. Kudo, A dual mesh method with adaptivity for stress-constrained topology optimization, Structural and Multidisciplinary Optimization 61 (2020) 749–762. doi:10.1007/s00158-019-02393-6.DOI
  8. [8]Y. Xie, T. Takikawa, S. Saito, O. Litany, S. Yan, N. Khan, F. Tombari, J. Tompkin, V. Sitzmann, S. Sridhar, Neural Fields in Visual Computing and Beyond, Computer Graphics Forum 41 (2022) 641–676. doi:10.1111/cgf.14505.DOI
  9. [9]L. Herrmann, O. Sigmund, V. M. Li, C. Vogl, S. Kollmannsberger, Neural Networks for Generating Better Local Optima in Topology Optimization, 2024. doi:10.48550/arXiv.2407.17957. arXiv:2407.17957.DOI
  10. [10]S. M. Sanu, A. M. Aragon, M. A. Bessa, Neural topology optimization: The good, the bad, and the ugly, 2024. doi:10.48550/arXiv.2407.13954. arXiv:2407.13954.DOI
  11. [11]Z. Zhang, Y. Li, W. Zhou, X. Chen, W. Yao, Y. Zhao, TONR: An exploration for a novel way combining neural network with topology optimization, Computer Methods in Applied Mechanics and Engineering 386 (2021) 114083. doi:10.1016/j.cma.2021.114083.DOI
  12. [12]N. Doosti, J. Panetta, V. Babaei, Topology Optimization via Frequency Tuning of Neural Design Representations, in: Symposium on Computational Fabrication, ACM, Virtual Event USA, 2021, pp. 1–9. doi:10.1145/3485114.3485124.DOI
  13. [13]T. Hospedales, A. Antoniou, P. Micaelli, A. Storkey, Meta-Learning in Neural Networks: A Survey, IEEE transactions on pattern analysis and machine intelligence 44 (2022) 5149–5169. doi:10.1109/TPAMI.2021.3079209.DOI
  14. [14]C. Finn, P. Abbeel, S. Levine, Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, 2017. doi:10.48550/arXiv.1703.03400. arXiv:1703.03400.DOI
  15. [15]A. Nichol, J. Achiam, J. Schulman, On First-Order Meta-Learning Algorithms, 2018. doi:10.48550/arXiv.1803.02999. arXiv:1803.02999.DOI
  16. [16]V. Sitzmann, E. R. Chan, R. Tucker, N. Snavely, G. Wetzstein, MetaSDF: Meta-learning Signed Distance Functions, 2020. arXiv:2006.09662.arxiv.org/abs/2006.09662
  17. [17]M. Tancik, B. Mildenhall, T. Wang, D. Schmidt, P. P. Srinivasan, J. T. Barron, R. Ng, Learned Initializations for Optimizing Coordinate-Based Neural Representations, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 2846–2855.arxiv.org/abs/2012.02189
  18. [18]J. Tack, S. Kim, S. Yu, J. Lee, J. Shin, J. R. Schwarz, Learning Large-scale Neural Fields via Context Pruned Meta-Learning, 2023. doi:10.48550/arXiv.2302.00617. arXiv:2302.00617.arxiv.org/abs/2302.00617
  19. [19]X. Huang, Z. Ye, H. Liu, B. Shi, Z. Wang, K. Yang, Y. Li, B. Weng, M. Wang, H. Chu, F. Yu, B. Hua, L. Chen, B. Dong, Meta-Auto-Decoder for Solving Parametric Partial Differential Equations, 2022. doi:10.48550/arXiv.2111.08823. arXiv:2111.08823.arxiv.org/abs/2111.08823
  20. [20]M. Penwarden, S. Zhe, A. Narayan, R. M. Kirby, A Metalearning Approach for Physics-Informed Neural Networks (PINNs): Application to Parameterized PDEs, 2023. doi:10.1016/j.jcp.2023.11191211912. arXiv:2110.13361.DOI
  21. [21]T. Qin, A. Beatson, D. Oktay, N. McGreivy, R. P. Adams, Meta-PDE: Learning to Solve PDEs Quickly Without a Mesh, 2022. doi:10.48550/arXiv.2211.01604. arXiv:2211.01604.DOI
  22. [22]Y. Yu, T. Hur, J. Jung, I. G. Jang, Deep learning for determining a near-optimal topological design without any iteration, Structural and Multidisciplinary Optimization 59 (2019) 787–799. doi:10.1007/s00158-018-2101-5.DOI
  23. [23]Z. Nie, T. Lin, H. Jiang, L. B. Kara, TopologyGAN: Topology Optimization Using Generative Adversarial Networks Based on Physical Fields Over the Initial Domain, Journal of Mechanical Design 143 (2021) 031715. doi:10.1115/1.4049533.DOI
  24. [24]A. H. Nobari, G. Giannone, L. Regenwetter, F. Ahmed, NITO: Neural Implicit Fields for Resolution-free Topology Optimization, 2024. doi:10.48550/arXiv.2402.05073. arXiv:2402.05073.DOI
  25. [25]A. Lutheran, S. Das, A. Tabarraei, Latent Space Diffusion for Topology Optimization, 2025. doi:10.48550/arXiv.2508.05624. arXiv:2508.05624.DOI
  26. [26]A. H. Nobari, L. Regenwetter, C. Picard, L. Han, F. Ahmed, Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization, 2025. doi:10.48550/arXiv.2510.23667. arXiv:2510.23667.DOI
  27. [27]I. Sosnovik, I. Oseledets, Neural networks for topology optimization, 2017. doi:10.48550/arXiv.1709.09578. arXiv:1709.09578.arxiv.org/abs/1709.09578
  28. [28]N. A. Kallioras, G. Kazakis, N. D. Lagaros, Accelerated topology optimization by means of deep learning, Structural and Multidisciplinary Optimization 62 (2020) 1185–1212. doi:10.1007/s00158-020-02545-z.DOI
  29. [29]J. Lim, K. Jung, Y. Jung, D.-N. Kim, Accelerating topology optimization using deep learning-based image super-resolution, Engineering Applications of Artificial Intelligence 133 (2024) 108370. doi:10.1016/j.engappai.2024.108370.DOI
  30. [30]R. V. Woldseth, N. Aage, J. A. Bærentzen, O. Sigmund, On the use of Artificial Neural Networks in Topology Optimisation, Structural and Multidisciplinary Optimization 65 (2022) 294. doi:10.1007/s00158-022-03347-1. arXiv:2208.02563.DOI
  31. [31]S. Shin, D. Shin, N. Kang, Topology optimization via machine learning and deep learning: A review, Journal of Computational Design and Engineering 10 (2023) 1736–1766. doi:10.1093/jcde/qwad072.DOI
  32. [32]S. Hoyer, J. Sohl-Dickstein, S. Greydanus, Neural reparameterization improves structural optimization, 2019. doi:10.48550/arXiv.1909.04240. arXiv:1909.04240.DOI
  33. [33]A. Chandrasekhar, K. Suresh, Approximate length scale filter in topology optimization using fourier enhanced neural networks, Computer-Aided Design 150 (2022) 103277. doi:10.1016/j.cad.2022.103277.DOI
  34. [34]B. Dupuis, A. Jacot, DNN-based topology optimisation: Spatial invariance and neural tangent kernel, in: M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, J. W. Vaughan (Eds.), Advances in Neural Information Processing Systems, volume 34, Curran Associates, Inc., 2021, pp. 27659–27669.arxiv.org/abs/2106.05710
  35. [35]M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-López, F. Pedregosa, J.-P. Vert, Efficient and Modular Implicit Differentiation, 2022. doi:10.48550/arXiv.2105.15183. arXiv:2105.15183.DOI
  36. [36]S. Pan, S. L. Brunton, J. N. Kutz, Neural Implicit Flow: A mesh-agnostic dimensionality reduction paradigm of spatio-temporal data, Journal of Machine Learning Research 24 (2023) 1–60.arxiv.org/abs/2204.03216
  37. [37]N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. A. Hamprecht, Y. Bengio, A. Courville, On the Spectral Bias of Neural Networks, 2019. doi:10.48550/arXiv.1806.08734. arXiv:1806.08734.DOI
  38. [38]H. Chen, A. Joglekar, L. Burak Kara, Topology Optimization Using Neural Networks With Conditioning Field Initialization for Improved Efficiency, Journal of Mechanical Design 146 (2023). doi:10.1115/1.4064131.DOI
  39. [39]J. Zehnder, Y. Li, S. Coros, B. Tomaszewski, NTopo: Mesh-free topology optimization using implicit neural representations, in: M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, J. W. Vaughan (Eds.), Advances in Neural Information Processing Systems, volume 34, Curran Associates, Inc., 2021, pp. 10368–10381.arxiv.org/abs/2102.10782
  40. [40]B. Bourdin, Filters in topology optimization, International Journal for Numerical Methods in Engineering 50 (2001) 2143–2158. doi:10.1002/nme.116.DOI
  41. [41]D. P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization, 2017. doi:10.48550/arXiv.1412.6980. arXiv:1412.6980.DOI
  42. [42]J. R. R. A. Martins, A. Ning, Engineering Design Optimization, Cambridge University Press, Cambridge, 2021.DOI
  43. [43]O. Sigmund, On benchmarking and good scientific practise in topology optimization, Structural and Multidisciplinary Optimization 65 (2022) 315. doi:10.1007/s00158-022-03427-2.DOI
  44. [44]F. Mazé, F. Ahmed, Diffusion Models Beat GANs on Topology Optimization, 2022. doi:10.48550/arXiv.2208.09591. arXiv:2208.09591.arxiv.org/abs/2208.09591
  45. [45]S. G. Johnson, The NLopt nonlinear-optimization package, 2007.
  46. [46]E. D. Dolan, J. J. Moré, Benchmarking Optimization Software with Performance Profiles, 2004. doi:10.48550/arXiv.cs/0102001. arXiv:cs/0102001.
  47. [47]S. Rojas-Labanda, M. Stolpe, Benchmarking optimization solvers for structural topology optimization, Structural and Multidisciplinary Optimization 52 (2015) 527–547. doi:10.1007/s00158-015-1250-z.
  48. [48]H. A. Eschenauer, V. V. Kobelev, A. Schumacher, Bubble method for topology and shape optimization of structures, Structural optimization 8 (1994) 42–51. doi:10.1007/BF01742933.DOI
  49. [49]V. Young, O. M. Querin, G. P. Steven, Y. M. Xie, 3D and multiple load case bidirectional evolutionary structural optimization (BESO), Structural optimization 18 (1999) 183–192. doi:10.1007/BF01195993.DOI
  50. [50]S. Yoo, N. Kang, DeepWheel: Generating a 3D Synthetic Wheel Dataset for Design and Performance Evaluation, 2025. doi:10.48550/arXiv.2504.11347. arXiv:2504.11347.arxiv.org/abs/2504.11347
  51. [51]Z. Li, F. Zhou, F. Chen, H. Li, Meta-SGD: Learning to Learn Quickly for Few-Shot Learning, 2017. doi:10.48550/arXiv.1707.09835. arXiv:1707.09835.DOI
  52. [52]M. E. Lynch, S. Sarkar, K. Maute, Machine Learning to Aid Tuning of Numerical Parameters in Topology Optimization, Journal of Mechanical Design 141 (2019). doi:10.1115/1.4044228.DOI
  53. [53]S. Yan, F. Wang, O. Sigmund, On the non-optimality of tree structures for heat conduction, International Journal of Heat and Mass Transfer 122 (2018) 660–680. doi:10.1016/j.ijheatmasstransfer.2018.01.114.DOI
  54. [54]M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, R. Ng, Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains, 2020. doi:10.48550/arXiv.2006.10739. arXiv:2006.10739.DOI

Paper details

vlnc:W_d9xpgs74
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Claude (Anthropic)

View statement
Looking for related papers…
Contents