Iterative Residual SVID for SIW Design

• The paper introduces a two-stage deep learning framework that combines a feedforward inverse model with an iterative residual correction network, reducing parameter estimation errors by over 50%.
• IRC-Net refines the initial SIW geometry predictions over five iterative corrections, achieving a test MSE of approximately 0.00146 and a significant speedup over full-wave simulations.
• The method reduces computational time from about 3 minutes per simulation to under 50 ms, enabling rapid prototyping of complex Ku-band microwave structures.

Iterative Residual SVID refers to a two-stage deep learning framework for high-precision inverse design of microwave structures, specifically Ku-band Substrate Integrated Waveguide (SIW) resonant components based on multimode resonators. The approach leverages an initial feedforward inverse model, followed by multiple passes of a residual correction network. This method is known as the Iterative Residual Correction Network (IRC-Net), which demonstrates significant gains in inverse modeling accuracy, generalization, and computational efficiency over traditional single-stage neural networks and conventional electromagnetic (EM) solvers (Mashayekhi et al., 11 May 2025).

1. Inverse Design Formulation

The inverse design framework addresses the mapping from high-dimensional electromagnetic (EM) response characteristics to the underlying geometric parameters of complex SIW resonators. The geometric design parameter vector is specified as

$$\mathbf{D}=(D_1,\,D_2,\,R_1,\,R_2,\,R_3,\,G)\in\mathbb R^6,$$

where each variable encodes a dimensional attribute of the multimode SIW structure. The associated EM response consists of the modulus of S-parameters $|S_{11}(f_i)|,\,|S_{21}(f_i)|$ sampled at 1,001 discrete frequencies spanning 12–18 GHz. The response vector is

$$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$

The primary inverse problem is, for a specified target response $\mathbf{S}_{\rm targ}$, to identify $\mathbf{D}$ for which the forward model (full-wave solver, denoted $\mathcal F$) returns an output close to $\mathbf{S}_{\rm targ}$. Formally,

$$\min_{\mathbf{D}}\;\bigl\|\mathcal F(\mathbf{D})\;-\;\mathbf{S}_{\rm targ}\bigr\|_2^2.$$

The problem is tackled by learning a neural inverse mapping $g:\mathbf{S}\to\mathbf{D}$, where $$g(\mathbf{S}_{\rm targ})\approx\mathbf{D}_{\rm true}$$.

2. Feedforward Inverse Model (FIM)

The FIM serves as the first stage, producing an initial estimate for the geometry given a target EM response. The FIM is a seven-layer densely connected neural network:

Layer	Input→Output	Activation	Dropout
1	2002→1500	ReLU
2	1500→1000	ReLU	0.1
3	1000→500	ReLU	0.1
4	500→250	ReLU	0.1
5	250→125	ReLU	0.1
6	125→64	ReLU	0.1
7	64→32	ReLU
Output	32→6	Linear

The model is trained to minimize the mean-squared error (MSE) loss

$|S_{11}(f_i)|,\,|S_{21}(f_i)|$0

The optimizer is Adam; early stopping is monitored on validation loss, with training conducted over approximately 200 epochs. On validation data, FIM achieves MSE $|S_{11}(f_i)|,\,|S_{21}(f_i)|$1 and MAE $|S_{11}(f_i)|,\,|S_{21}(f_i)|$2 (Mashayekhi et al., 11 May 2025).

3. Iterative Residual Correction Network (IRC-Net)

IRC-Net implements a correction mechanism atop the FIM's coarse predictions. At each correction iteration $|S_{11}(f_i)|,\,|S_{21}(f_i)|$3 ($|S_{11}(f_i)|,\,|S_{21}(f_i)|$4 with $|S_{11}(f_i)|,\,|S_{21}(f_i)|$5 fixed), the residual between the true geometry and the current prediction is computed:

$|S_{11}(f_i)|,\,|S_{21}(f_i)|$6

A small residual network $|S_{11}(f_i)|,\,|S_{21}(f_i)|$7 predicts the parameter update $|S_{11}(f_i)|,\,|S_{21}(f_i)|$8, with

$|S_{11}(f_i)|,\,|S_{21}(f_i)|$9

The residual network structure per iteration is:

Layer	Input→Output	Activation
Dense Layer 1	6→64	LeakyReLU
Dense Layer 2	64→64	LeakyReLU
Output	64→6	Linear

The only skip connection is the outer sum with the previous estimate ($$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$0). After $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$1 iterations, the output $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$2 is compared against the ground-truth using final parameter MSE:

$$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$3

4. Data Generation and Training Protocol

The dataset consists of 8,721 EM-simulated samples, with geometry parameters constrained by $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$4 and a geometric feasibility inequality involving all six parameters. Parameter grids are $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$5 mm (step 0.5 mm), $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$6 mm (step 0.2 mm), and $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$7 mm. Each sample requires approximately 3 minutes of full-wave simulation (CST). The data split is 85% training and 15% test; all features and targets are standardized. Dropout of 10% is applied throughout. Batch size is typically 32, with learning rate approximately $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$8, and early stopping on validation MSE.

5. Performance Evaluation

The IRC-Net approach provides substantial improvements in both accuracy and efficiency compared to single-stage models. After five correction steps, IRC-Net achieves test MSE $$\mathbf{S} = \bigl\{|S_{11}(f_i)|,\;|S_{21}(f_i)|\bigr\}_{i=1}^{1001}\in\mathbb R^{2002}.$$9 and MAE $\mathbf{S}_{\rm targ}$0, which constitutes a reduction in parameter estimation error by over 50% relative to the FIM baseline. Error histograms (see (Mashayekhi et al., 11 May 2025), Figs. 12–13) and error-versus-iteration plots (Fig. 11) demonstrate that most sample errors cluster near zero and converge quickly, with the sharpest drop after the first correction.

Experimental validation includes the design, fabrication, and measurement of both three- and four-resonance SIW structures:

• For three-resonator SIW, trained and fabricated devices show strong agreement between simulated and measured $\mathbf{S}_{\rm targ}$1 and $\mathbf{S}_{\rm targ}$2, confirming the method's predictive fidelity.
• For a four-resonator topology, IRC-Net, trained solely on three-resonator data, is applied to predict the geometry based on a target four-pole response. Fabrication and measurement again demonstrate close response matching (see (Mashayekhi et al., 11 May 2025), Fig. 15).

In terms of computational cost, conventional EM simulations require $\mathbf{S}_{\rm targ}$33 minutes per sample, while IRC-Net inference completes in under 50 ms on a CPU, representing a speedup of four orders of magnitude.

6. Generalization and Methodological Insights

The coarse-to-fine architecture (FIM followed by iterative correction) is broadly applicable to inverse mapping problems where a decent initial guess is computable. The iterative nature allows the residual network to focus on modeling small, high-frequency errors, thus improving generalization and fine-tuning. The lightweight residual blocks can be replaced or augmented with deeper networks or physics-informed layers in other contexts.

A plausible implication is that this strategy may be adapted for multi-objective inverse design tasks, by incorporating additional loss terms representing fabrication tolerance, coupling factor penalties, or other practical constraints. The framework is readily transferable to design challenges involving filters, metamaterials, or antennas, contingent on dataset retraining.

7. Broader Impact and Extensions

Iterative residual correction provides a robust, general strategy for high-precision inverse electromagnetic design. The method demonstrates both statistical and physical accuracy, with successful predictions validated by fabricated hardware. Its high computational efficiency makes it significantly advantageous for rapid prototyping and optimization workflows in electromagnetic engineering and beyond. This suggests applicability across a broad range of high-dimensional inverse problems where classical direct inversion is intractable or ill-posed (Mashayekhi et al., 11 May 2025).