Surrogate-Guided Inverse Design

• Surrogate-guided inverse design is a class of methods where learned models replace expensive full-scale simulations to guide design optimization.
• It integrates techniques like gradient-based optimization, Bayesian inference, and mixed-integer reformulations applicable to nanophotonics, CFD, and materials design.
• The approach substantially reduces computational cost while maintaining design fidelity, though challenges remain in handling domain shifts and surrogate accuracy.

Searching arXiv for papers on surrogate-guided inverse design across photonics, fluids, materials, and Bayesian inference. Surrogate-guided inverse design is a class of inverse-design methodologies in which an expensive forward model—typically a full-wave electromagnetic solver, a CFD code, a homogenization engine, or another PDE-constrained simulator—is replaced, complemented, or constrained by a learned surrogate during search, optimization, or posterior inference. In the reported literature, the surrogate may act as a differentiable replacement for the forward operator, as a guidance model inside diffusion or flow-matching samplers, as a compact matrix-space proxy for a photonic device, or as a piecewise-linear neural predictor that admits exact mixed-integer reformulation (Augenstein et al., 2023, Yang et al., 9 Dec 2025, Muda et al., 23 Apr 2026, Tiwari et al., 25 May 2026, Ansari et al., 2021). The common purpose is to reduce the cost of repeated forward evaluations while preserving enough fidelity to steer designs toward target functionality, quantify uncertainty, or enforce manufacturability and prior structure.

1. Scope and defining characteristics

Surrogate-guided inverse design is defined most explicitly in electromagnetic nanophotonics as the replacement of “expensive full-wave electromagnetic solves inside an optimization loop with a differentiable, data-driven surrogate that predicts fields with sufficient accuracy to steer the design” (Augenstein et al., 2023). Closely related formulations appear in Bayesian CFD, where a neural operator is “embedded directly within the MCMC inference loop while preserving the likelihood model, priors, and sampling configuration,” and in photonic neural networks, where a surrogate workflow “decouples task learning from electromagnetic realization” by separating matrix-space learning from full-wave operator transfer (Tiwari et al., 25 May 2026, Muda et al., 23 Apr 2026).

The term also covers generative settings. In aerodynamic design, Dflow-SUR uses a surrogate physical loss together with differentiation through a flow-matching generative model, explicitly separating physical-loss optimization from the denoising or transport dynamics (Yang et al., 9 Dec 2025). In metasurface design, a conditional diffusion model is regularized by a pretrained surrogate EM simulator so that generated geometries adhere to target reflection spectra (Joy et al., 19 May 2026). In subsurface inversion, SURGIN combines a score-based generative prior with a differentiable U-FNO surrogate so that unseen observations can guide posterior sampling without retraining (Feng et al., 16 Sep 2025). In RNA design, structural self-consistency metrics computed from folding predictors function as surrogate rewards for reinforcement-learning fine-tuning, replacing native sequence recovery as the operative design signal (Hu et al., 18 Feb 2026).

A recurrent misconception is that surrogate-guided inverse design is synonymous with a single optimization paradigm. The cited work instead spans gradient-based optimization through differentiable surrogates, Bayesian posterior sampling with neural operators, mixed-integer reformulations of piecewise-linear neural surrogates, greedy surrogate-based search, variational annealing in latent spaces, and reinforcement learning guided by surrogate rewards (Augenstein et al., 2023, Tiwari et al., 25 May 2026, Ansari et al., 2021, Grbcic et al., 2024, Bezick et al., 2024, Hu et al., 18 Feb 2026). This suggests that the unifying feature is not the optimizer, but the role of the surrogate as the computationally tractable stand-in, guide, or prior-aware intermediary between design variables and physical objectives.

2. Mathematical structure of the inverse problem

A recurring formulation is a forward map from design variables to fields, observables, or task metrics, followed by an inverse objective that measures deviation from a target. In electromagnetic scattering, the surrogate learns the operator

$$\mathcal{G}: \varepsilon_r(\mathbf{x}) \mapsto \mathbf{E}(\mathbf{x}),$$

with the governing frequency-domain Maxwell system

$$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$

and inverse design optimizes functionals of the predicted fields, such as

$$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$

for single-spot or four-spot focusing (Augenstein et al., 2023).

In Bayesian inverse design for quasi-one-dimensional nozzle flow, the observation model is written as

$y = H(G(\theta)) + \epsilon,$

with Gaussian likelihood

$$p(y \mid \theta, \sigma^2) \propto \exp \left( -\frac{1}{2\sigma^2} \sum_{j=1}^{N_{\mathrm{obs}}} \big( y_j - H_j(G(\theta)) \big)^2 \right),$$

and surrogate-guided inference replaces the CFD operator inside $H(G(\theta))$ by a differentiable DeepONet while leaving the likelihood, priors, and NUTS configuration unchanged (Tiwari et al., 25 May 2026). SURGIN adopts an analogous Bayesian decomposition in function space,

$$p(\theta \mid y) \propto p(y \mid \theta) p(\theta),$$

but evaluates the likelihood gradient through a differentiable U-FNO surrogate and a score-based generative prior (Feng et al., 16 Sep 2025).

Other formulations operate on surrogate losses rather than field residuals. In photonic neural networks, the realization stage minimizes a scattering-operator residual of the form

$$L = \|T_{\mathrm{device}} - T_{\mathrm{target}}\|_F^2 + \lambda \|R_{\mathrm{device}}\|_F^2 + \Omega_{\mathrm{reg}},$$

which is batch-free once the target operator is fixed (Muda et al., 23 Apr 2026). In Dflow-SUR, the inverse-design loss may be a target-matching quadratic such as

$$L_{\mathrm{phys}}(x) = \alpha [C_L(x) - C_L^{\mathrm{target}}]^2 + \beta [C_D(x) - C_D^{\mathrm{target}}]^2 + \sum_{j=1}^J \lambda_j \,\mathrm{penalty}_j(x),$$

or a lift-to-drag objective

$$L_{\mathrm{phys}}(x) = -\gamma \frac{C_L(x)}{C_D(x)} + \sum_{j=1}^J \lambda_j \,\mathrm{penalty}_j(x),$$

with gradients propagated back to the source noise of the generative model (Yang et al., 9 Dec 2025).

A more explicitly prior-constrained formulation appears in Deep Physics Prior, where the design is parameterized as $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$0 and optimization is performed in latent space:

$$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$1

Here the surrogate forward operator $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$2 and the auxiliary generative prior $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$3 jointly define the admissible search manifold (Yang et al., 28 Apr 2025). Mixed Integer Neural Inverse Design takes a different route: when the surrogate is piecewise linear, inverse design can be posed exactly as a mixed-integer linear program over the design variables and neural-network activation binaries, enabling globally optimal or near-optimal solutions under bounded domains (Ansari et al., 2021).

3. Surrogate architectures and design parameterizations

The surrogate families used in the literature are heterogeneous and strongly tied to the governing physics. Neural operators are prominent in PDE-governed settings. The electromagnetic inverse-design work on free-form scatterers uses a modified Fourier Neural Operator with three stated changes relative to vanilla FNO: zero-padding in real space rather than coordinate-feature lifting, GELU activations, and spectral truncation tuned to electromagnetic wave content (Augenstein et al., 2023). Bayesian nozzle inversion employs a DeepONet with a branch network for the geometry function and a trunk network for spatial coordinates, while SURGIN uses a U-Net enhanced Fourier Neural Operator so that multiscale spatial structure and differentiability are preserved during posterior guidance (Tiwari et al., 25 May 2026, Feng et al., 16 Sep 2025).

Physics-guided encoder–decoder surrogates form another major branch. The Theory-guided Auto-Encoder embeds finite-difference residuals of the governing PDE, together with boundary and initial conditions, directly into the loss of a convolutional encoder–decoder surrogate for transient subsurface flow (Wang et al., 2020). In photonic explainability work, a lightweight CNN is trained on SPINS-B generated wavelength demultiplexers so that Integrated Gradients can be computed on the surrogate instead of the full solver (Park et al., 25 Oct 2025). In optical waveform design for nonlinear electromagnetic dynamics, a 1D CNN surrogate maps a 301-sample temporal waveform to a scalar emittance objective, with positivity and fixed pulse energy enforced during the inverse loop (Zhang et al., 12 Mar 2026).

Classical surrogates remain competitive in lower-dimensional or tabular regimes. Random Forest regressors are used in greedy laser-parameter search for photonic surfaces and for real-time inverse analysis of auxetic metamaterials (Grbcic et al., 2024, Danesh et al., 2024). Gradient-enhanced Gaussian processes are used for inverse problems with adaptive design-of-experiments, while Gaussian process regression and Chebyshev polynomial interpolation are compared directly for drying-induced assembly of colloidal films (Semler et al., 2024, Kundu et al., 12 Sep 2025). PearSAN adopts a pseudo-Boolean surrogate in a discrete latent space and trains it with a Pearson-correlational loss so that the surrogate energy is monotonic with the true figure of merit rather than absolutely calibrated (Bezick et al., 2024).

Design parameterization is equally consequential. In electromagnetic free-form design, direct voxel optimization is explicitly avoided because continuous $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$4 values induce distribution drift; instead, geometry is parameterized by the latent vector of a 3D convolutional VAE whose decoder yields approximately binary structures (Augenstein et al., 2023). In Dflow-SUR, airfoils use 16 CST parameters and wings use a compact modal parameterization, while the generative prior is a continuous normalizing flow defined by flow matching (Yang et al., 9 Dec 2025). In photonic neural networks, the stage-one surrogate design variable is a passive complex matrix

$$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$5

with bounded singular values, later transferred to a fabrication-aware freeform device (Muda et al., 23 Apr 2026). In spinodoid metamaterials, the surrogate exploits permutation equivariance with respect to $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$6 and embeds orthorhombic symmetry, isotropy limits, and positive semidefiniteness directly into the architecture, which is the stated basis for achieving accurate surrogate performance with only 75 data points (Rosenkranz et al., 6 May 2025).

4. Inverse-design algorithms and guidance mechanisms

The most direct surrogate-guided strategy is end-to-end differentiation through a surrogate forward map. In three-dimensional nanophotonic design, gradients are computed by automatic differentiation through the composition

$$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$7

with the FNO providing $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$8 and the VAE decoder providing $$\nabla \times (\mu^{-1} \nabla \times \mathbf{E}) - \omega^2 \varepsilon(\mathbf{x}) \mathbf{E} = i \omega \mathbf{J}(\mathbf{x}),$$9 (Augenstein et al., 2023). DPP follows the same principle in latent space, but with a learned prior manifold:

$$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$0

so that first-order optimization remains within the range of the pretrained generator (Yang et al., 28 Apr 2025).

A second family of methods uses surrogates to guide probabilistic inference. In the nozzle study, the forward DeepONet is frozen and inserted into NUTS; all gradients required by the sampler pass through the neural operator unchanged, and the posterior definition itself is not altered (Tiwari et al., 25 May 2026). SURGIN is more explicitly generative: a score-based prior over geological fields is combined with a surrogate-derived likelihood gradient, yielding a guided reverse SDE in which the posterior score decomposes into prior and likelihood contributions (Feng et al., 16 Sep 2025). IP-SUR and CSQ address a related problem at the sequential-design level by adaptively choosing where to evaluate an expensive forward model so that a Gaussian-process surrogate becomes most useful for Bayesian inverse problems; IP-SUR specifically minimizes a posterior-weighted integrated uncertainty functional and is reported with an almost sure convergence guarantee (Lartaud et al., 2024).

Generative inverse design introduces additional guidance mechanisms. Dflow-SUR samples a terminal design $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$1 by integrating the learned flow, evaluates $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$2 only at the terminal state, and then updates the source noise via

$$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$3

thereby decoupling the number of optimization steps from the ODE discretization (Yang et al., 9 Dec 2025). Metasurface diffusion instead imposes a surrogate spectrum loss during training,

$$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$4

so that samples are already biased toward spectral validity before any optional test-time guidance (Joy et al., 19 May 2026). PearSAN uses a discrete latent-space annealing procedure: a pseudo-Boolean surrogate is trained to be Pearson-correlated with the figure of merit, and a recurrent variational annealer minimizes the surrogate free energy in latent space (Bezick et al., 2024). RIDER replaces direct differentiability entirely with policy-gradient optimization over a diffusion policy, using structural rewards such as GDT_TS, RMSD, TM-score, and a composite reward computed from a folding oracle (Hu et al., 18 Feb 2026).

Not all surrogate-guided inverse design is gradient-based or stochastic. ALPS performs greedy surrogate-based search by repeatedly training a Random Forest surrogate, sampling a Latin-hypercube candidate set, selecting candidates with lowest surrogate-predicted RMSE to the target spectrum, and augmenting the dataset with newly evaluated points (Grbcic et al., 2024). Mixed Integer Neural Inverse Design exploits the piecewise-linear structure of ReLU surrogates to formulate inverse design exactly as a mixed-integer linear program, which is especially useful when combinatorial constraints such as material selection are central (Ansari et al., 2021). This suggests that surrogate guidance can operate through exact optimization structure as well as through backpropagation.

5. Performance across application domains

Reported gains are substantial but domain-specific. In free-form electromagnetic inverse design, the modified FNO trained on 8,192 three-dimensional samples achieved a median normalized $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$5 on 400 test geometries for complex vector fields, while inference took approximately $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$6 seconds on an A100 compared with approximately $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$7 minutes per full-wave FDTD sample. A 64-start, 300-iteration inverse-design campaign completed in approximately $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$8 minutes wall time on 32 GPUs, whereas a traditional adjoint-based design was estimated at approximately $$J(\boldsymbol{z}) = \sum_{\mathbf{r} \in \mathcal{D}} |\mathbf{E}(\mathbf{r}; \boldsymbol{z})|^2$$9 hours per single run at $y = H(G(\theta)) + \epsilon,$0 minutes per solve (Augenstein et al., 2023).

In aerodynamic generative design, Dflow-SUR reports a reduction in physical loss by four orders of magnitude and a 74% cut in wall-clock time on the airfoil case relative to the strongest energy-based baseline. For wings, the mean lift-to-drag ratio was 21.18 versus 18.40 for Latin-hypercube sampling and 19.84 for the energy-based method, with the standard deviation reduced to 0.70 from 1.46 and 1.02, respectively (Yang et al., 9 Dec 2025). Photonic neural networks show a different type of gain: by separating matrix-space learning from device realization, the realization stage removes minibatch dependence from the full-wave loop, reducing simulation counts by 1–3 orders of magnitude. On MedMNIST, the realized all-optical classifier reached 98.16% test accuracy versus 98.75% for the surrogate, within 0.59 percentage points after only 20 adjoint epochs (Muda et al., 23 Apr 2026).

In uncertainty-aware CFD, replacing the quasi-1D nozzle solver with a DeepONet reduced NUTS runtime from approximately 42 minutes per run to under one second, corresponding to a speedup of order $y = H(G(\theta)) + \epsilon,$1, while preserving posterior mean structure and credible-interval contraction trends across observation regimes (Tiwari et al., 25 May 2026). In metasurface absorber design, the physics-guided conditional diffusion framework achieved an average spectral MSE of approximately 0.0006, band alignment accuracy of approximately 0.958, and approximately 30 seconds per design, whereas the conventional approach was described as taking several months under comparable computational resources (Joy et al., 19 May 2026).

Material and mechanics applications exhibit comparable acceleration. The FFT-based auxetic metamaterial framework uses Random Forest surrogates and a brute-force inverse search that remains below 1 second for 20k evaluations; the rectangular-void case study achieved FEM-validated $y = H(G(\theta)) + \epsilon,$2 MPa, $y = H(G(\theta)) + \epsilon,$3 MPa, and $y = H(G(\theta)) + \epsilon,$4 for a target of $y = H(G(\theta)) + \epsilon,$5 MPa and $y = H(G(\theta)) + \epsilon,$6 MPa (Danesh et al., 2024). The spinodoid metamaterial surrogate reaches reliable multi-objective inverse design with only 75 data points, far below the several thousands or hundreds of thousands cited for earlier neural-network-based approaches (Rosenkranz et al., 6 May 2025). PearSAN reports a maximum thermophotovoltaic design efficiency of 97.02% and 0.0033 hours per 100 designs, while DPP reports 0.01 EPE violations at 0.4 seconds throughput in inverse lithography, compared with 0.21 violations at 4.0 seconds for the numerical baseline (Bezick et al., 2024, Yang et al., 28 Apr 2025).

6. Validation, limitations, and methodological tensions

The central reliability issue is distributional validity. The electromagnetic FNO study is explicit that the surrogate is specialized to a single frequency, a single illumination condition, binary $y = H(G(\theta)) + \epsilon,$7 with $y = H(G(\theta)) + \epsilon,$8, and a fixed domain and resolution; performance degrades outside this manifold, and direct voxel optimization without a binary geometry prior produces unreliable predictions (Augenstein et al., 2023). The metasurface diffusion framework notes analogous risks under surrogate domain shift and out-of-distribution target spectra (Joy et al., 19 May 2026). The Bayesian nozzle study observes that posterior fidelity is preserved when the training distribution covers the posterior support, but severe shocks and dense observations can expose localized surrogate error near discontinuities (Tiwari et al., 25 May 2026). DPP formalizes the same concern through an explicit error decomposition and argues that unconstrained first-order optimization can exploit surrogate error unless the search is restricted to the range of a pretrained prior generator (Yang et al., 28 Apr 2025).

Validation strategies therefore recur. Final photonic devices are re-evaluated by full-wave FDTD or FEM, promising candidates are screened by the true solver, and Bayesian studies compare surrogate-based posteriors against high-fidelity references under unchanged priors and likelihoods (Augenstein et al., 2023, Tiwari et al., 25 May 2026, Danesh et al., 2024). Dflow-SUR supplements performance metrics with uncertainty estimation via Monte Carlo dropout and argues that optimizing at denoised terminal samples avoids the high surrogate uncertainty of intermediate states (Yang et al., 9 Dec 2025). SURGIN quantifies uncertainty through posterior ensembles, reporting KL divergence for geological fields and SSIM for flow states under sparse wells, super-resolution, and inpainting scenarios (Feng et al., 16 Sep 2025).

A second tension concerns what counts as an adequate surrogate objective. RIDER shows that native sequence recovery is a weak surrogate for RNA three-dimensional fidelity: at roughly 50% NSR, folded designs can range from near-zero GDT_TS to 0.9, and high NSR does not guarantee the correct fold (Hu et al., 18 Feb 2026). The photonic interpretability pipeline makes a different point: Integrated Gradients explains the CNN surrogate, not the Maxwell solver itself, so attribution quality depends on surrogate fidelity and baseline choice (Park et al., 25 Oct 2025). PearSAN likewise argues that absolute value-matching losses are often inferior to monotone, rank-aligned surrogate objectives in large latent spaces (Bezick et al., 2024). These cases suggest that surrogate-guided inverse design is not only about approximating a forward map; it is also about choosing a surrogate objective whose inductive bias is aligned with the true design criterion.

Across the cited work, several research directions recur without yet being standardized. Active learning is repeatedly proposed as a remedy for edge-of-distribution failures (Augenstein et al., 2023, Kundu et al., 12 Sep 2025). Physics-informed regularization or explicit residual penalties are presented as compatible with operator learners, though not always used in the reported experiments (Augenstein et al., 2023, Wang et al., 2020). Uncertainty-aware design remains unevenly developed: some systems provide posterior ensembles or Monte Carlo dropout, whereas others rely primarily on held-out accuracy and final high-fidelity validation (Yang et al., 9 Dec 2025, Tiwari et al., 25 May 2026, Feng et al., 16 Sep 2025). A plausible implication is that the field is converging on a common architecture—differentiable surrogate, geometry or prior manifold, and selective recourse to the true simulator—but not yet on a single notion of trustworthiness or optimal guidance.