betaFeedback

WorksEngineering

Multi-Label Classification with Constrained Search for Inverse Design of a 16x16 Reconfigurable Pixel Antenna

Abstract

In this paper, an inverse design method is presented for a 16x16 reconfigurable pixel antenna with 480 binary switches, giving 2480 possible patterns, far beyond what full-wave simulation can explore. Given a target frequency between 1 and 6 GHz, the method searches for a switch configuration whose simulated resonance lies close to the target. A forward surrogate model is trained on 22,152 full- wave simulated patterns to predict where a given switch pattern resonates. The forward task is formulated as multi-label classification over 100 frequency bins rather than reflection-coefficient regression, because a curve-fitting objective is dominated by the flat baseline around the narrow resonances of this structure. The inverse design is a simulated annealing search over the switch state, with a connectivity constraint maintained at every step so that every pattern generated is a valid antenna. The proposed methodology combines resonance-focused surrogate modelling with connectivity-constrained discrete optimization, enabling candidate generation without full-wave solver calls during the search. The pipeline is verified in CST Microwave Studio by simulating 220 predicted candidates, ten for each of 22 target frequencies. For each request, the best of the ten candidates reached a median frequency error of 23 MHz across the 22 targets, compared with 35 MHz for a solver-guided search using 200 simulations for each target. The proposed methodology therefore enables automatic switch configuration to achieve a requested resonant frequency, eliminating solver calls during the search and manual redesign of the antenna geometry.

Introduction

Reconfigurable planar antennas have become central to modern wireless communication systems, where a single low-cost structure is expected to adapt its behaviour on demand [1, 2]. This adaptability is achieved by tuning elements such as switches, PIN diodes, or varactors, which allow the operating frequency, radiation pattern, or polarization of the antenna to be changed after fabrication [3, 4]. In one widely used approach these switches are placed between a grid of small metallic elements, known as pixels, so that the connections between pixels are changed and the radiating structure is reshaped when the switches are turned on or off. This class of structures, known as pixel antennas, is valued for its high reconfigurability, low cost, and compact form factor [2, 5]. By setting the state of every pixel connection, the current distribution, impedance match, and resonant behaviour of the antenna can be engineered, which makes pixel antennas a flexible platform for programmable wireless systems [6, 7].

The central problem in using such a board is inverse design. The user specifies a desired electromagnetic response, for example a resonant frequency, and the design task is to find a switch pattern that produces it. With NN binary bridges the design space contains 2N2^N patterns. For the antenna considered in this paper N=480N = 480, so the design space has approximately 3.1×101443.1 \times10^{144} elements. Direct enumeration is not possible. The proposed method is shown in Figure 1. A target frequency is specified, a switch pattern is found by a search guided by a trained surrogate model, and the result is verified by full-wave simulation.

Overview of the proposed inverse design method

Figure 1. Overview of the proposed inverse design method: a target resonant frequency is specified, the inverse search produces a switch pattern, and the design is verified in CST Microwave Studio.

Two families of methods have been used for related problems. In topology optimization the design is treated as a continuous density field and gradients of the electromagnetic objective with respect to that field are used [8], [9]. This approach is well established in mechanics and has been extended to nanophotonics and antenna design [10]. In genetic algorithms and other evolutionary methods the design is treated as a discrete vector and explored by mutation and selection [4], [11]. Both families share a structural difficulty: every candidate evaluation requires a full-wave electromagnetic simulation that takes several minutes on a workstation. The number of evaluations a typical budget allows is therefore in the hundreds or low thousands, and the resulting designs are often small or symmetric to keep the search tractable. Machine learning has been used to remove this evaluation bottleneck. A surrogate model is trained on a moderate number of full-wave simulations and is then used in place of the solver during the search [10], [12]–[15]. Once trained, a candidate design is evaluated by the surrogate in milliseconds. The forward surrogate can either be inverted directly by gradient-based optimization on the model input [16], [17] or used as a scoring function inside a discrete search such as a genetic algorithm or simulated annealing. The first approach is fast, but a smooth input space is required; the second is slower, but applicable to non-smooth or constraint-heavy designs. For pixel antennas, where the design variables are binary, the evaluation cost has been reduced in two ways. In the first, every switch location is modelled as an internal port, so that one multiport simulation allows any switch state to be evaluated by circuit analysis. This model has been combined with genetic algorithms [18] and with successive exhaustive Boolean optimization, demonstrated on a 5×55 \times5 pixel antenna with 44 possible connections [19]. Every evaluation still solves a network with one port per switch at each frequency point, so its cost rises steeply with the size of the board. In the second, an ensemble-learning surrogate guides an evolutionary search for digitally coded antennas of more than a thousand pixels [20]. Each run produces one fixed antenna for one specification, so a reconfigurable board would need a new optimisation for every requested state.

Within antenna design specifically, machine learning has been applied to the inverse problem in several forms. Tandem networks have been used to design multiband antennas by training a forward model and an inverse model together [21], an inverse artificial neural network has been used to map a requested response back to geometric parameters under multiple objectives [22], and deep networks have been used to generate microstrip antenna geometries from a specified response [23]. In these methods, the geometry is assumed to be defined by a small number of continuous parameters, such as lengths, widths, and slot positions, which enables direct regression of these parameters using an inverse network. That assumption does not hold for the structure considered here, where the design is a 480-element binary vector with no continuous parametrisation and where a single bit change can move a resonance by hundreds of megahertz.

Direct regression of the reflection-coefficient curve from the switch pattern has been the standard formulation in much of this literature, including pixel antennas trained with a loss weighted toward the resonance [24], but it is a poor match for the narrow resonances of the board considered here.

The geometry of a pixel antenna cannot be described by a few shape parameters, which is what makes its response hard to model in the first place [25]. More importantly, the resonances of these structures are narrowband. A deep dip in the reflection coefficient occupies only a small part of the frequency axis, while the response stays close to its baseline everywhere else. Two effects follow from this. First, neural networks generally approximate smooth, slowly varying functions more readily than sharp local features, making narrow dips particularly difficult to reproduce accurately [26]. Second, when the quantity of interest is rare relative to the baseline, a root-mean-square error is dominated by many baseline points, and standard error metrics are reduced more by fitting the common case than the rare one [27]. A model can therefore drive the training error down by predicting a nearly flat curve that follows the baseline and misses the resonances entirely, which is the behaviour observed when the full curve is regressed on the data used here. This makes root-mean-square error an unreliable training objective for this problem, motivating a formulation in which the presence of a resonance, rather than the exact curve value, becomes the quantity the model predicts.

In this paper, the inverse design problem is addressed for a specific 16×1616 \times16 reconfigurable pixel antenna with 480 binary switches. The approach has three components:

  1. The forward surrogate task is reformulated from scattering-parameter regression to multi-label classification over a discretized frequency axis. Each pattern is labelled by a 100-bin vector of Gaussian soft labels wide enough to span several bins, so that a narrow resonance still produces a usable training signal.

  1. The inverse search is a simulated annealing chain over the 480-bit switch vector. The connectivity constraint is maintained at every step, so every state generated during the search corresponds to a valid antenna. By applying a soft penalty on the number of active patches, the search is constrained to the region of input space where the forward surrogate has good training coverage.

  1. The complete pipeline is verified by full-wave simulation of 220 generated candidates spanning 22 target frequencies, rather than by reporting forward-model

self-evaluation only.

Together, these components address both limitations of the earlier pixel-antenna methods. On a board with 480 switches, more than ten times the connections in [19], a candidate state is evaluated in about 3 ms through a single forward pass of the surrogate rather than a network solve over every switch. The model is trained once, and each new frequency request is then answered by a 2.2-minute search with no solver calls.

The remainder of this paper is organized as follows: The reconfigurable pixel antenna and its switching topology are described in § 2. The pattern generator, simulation protocol, and dataset statistics are detailed in § 2.1. The forward surrogate model and its training procedure are outlined in Section IV, followed by the inverse search strategy in § 4. In § 5, the performance metrics of the forward model, full-wave verification of the complete pipeline, and a comparison against a solver-guided search are presented.

Furthermore, an extension of the method to two-port filter operation and the limitations of the present work are addressed in § 7, followed by the conclusion in § 8.

Reconfigurable Pixel Antenna

The structure is a 16×1616 \times16 lattice of square copper patches, each 3×33 \times3 mm and 35 μm35\ \mu\mathrm{m} thick, modelled on a Rogers RO4730G3 substrate of thickness 1.524 mm, relative permittivity εr=2.98\varepsilon_r = 2.98, and loss tangent tan⁡δ=0.0029\tan\delta= 0.0029. Adjacent patches are separated by a 3.4 mm gap, and the lattice spans 99×9999 \times99 mm edge to edge. A continuous copper ground plane of dimensions 130×115130 \times115 mm underlies the structure. The board dimensions match the band of interest. To first order, a conducting path over the ground plane resonates when its length is close to half a guided wavelength, λg/2=c/(2fεeff)\lambda_g/2 = c/(2f\sqrt{\varepsilon_{\mathrm{eff}}}), where cc is the speed of light and εeff\varepsilon_{\mathrm{eff}} the effective permittivity [28], [29]. For a 3 mm strip on this substrate, W/h=1.97W/h = 1.97, and the quasi-static microstrip expression gives εeff≈2.36\varepsilon_{\mathrm{eff}} \approx2.36 [28]. Half a guided wavelength is then about 98 mm at 1 GHz and 16 mm at 6 GHz. A path of two or three patch pitches therefore resonates near the upper edge of the band, while a resonance near the lower edge needs a path as long as the full 99 mm width of the lattice. Figure 2 shows the top-down view and the substrate stack.

Top-down and side view of the 16 × 16 reconfigurable pixel antenna

Figure 2. Top-down and side view of the 16×1616 \times16 reconfigurable pixel antenna. A representative pattern is shown with shaded patches and active bridges. Port P1 is fixed at column 1, row 8 in 1-based indexing.

A switchable copper bridge of dimensions 1×3.41 \times3.4 mm is placed in each gap between two adjacent patches, modelled as ideal copper when on and as an open circuit when off. Treating the connections as ideal during design follows earlier pixel-antenna optimisation, where connections were represented as hardwires [19], and it isolates the effect of the switch pattern from that of any particular switch technology. The effect of practical switches is discussed in § 7. Each patch is connected to the patch below by a vertical bridge, giving 15 bridges per column across 16 columns, for 240 vertical bridges. The full switch state is therefore a 480-element binary vector, and the design space has 2480≈3.1×101442^{480} \approx3.1 \times10^{144} elements.

For numerical processing, the switch state is stored as a tensor of shape (2,16,16)(2,16,16), where channel 0 holds the horizontal bridges and channel 1 holds the vertical bridges. The horizontal bridges naturally form a 15×1615 \times16 array and the vertical bridges a 16×1516 \times15 array, so the two sets do not share the same shape.

Each is padded to a full 16×1616 \times16 grid so that both can be stored together as two channels and each bridge stays aligned with the patch it belongs to. The cost of this alignment is one unused column in channel 0 and one unused row in channel 1, which are zero for every pattern. A single port, P1, is connected to the patch at position (1,8)(1,8) in 1-based indexing, which is the leftmost column, eighth row. The port consists of a tapered microstrip feed of 16 mm length, connected to a waveguide port at its outer edge, with the taper providing an impedance match between the port boundary and the patch interface near 3.5 GHz. The active bridges form a single connected metallic region that contains P1, and this region radiates and produces one or more resonant dips in S11S_{11}. Every pattern used in this work is a single connected component rooted at P1, so the reflection coefficient at P1 is the only observable utilized by the design method.

Dataset Generation

The forward model is trained on switch patterns paired with their simulated reflection coefficients, so a dataset has to be built by simulation before any model can be fitted. Which patterns enter that dataset is itself a design decision, since the 24802^{480} configuration space cannot be sampled exhaustively and every candidate costs a full-wave simulation. Training patterns are therefore produced by a stochastic generator in which one of eight pattern families is selected with equal probability. The families are a square region grown around P1, a rectangular region, a ring, a split-ring resonator, a small patch cluster with thin branches, a fractal, a breadth-first random blob from P1, and a hybrid combining a square base with breadth-first growth. Every family is anchored at P1. Each candidate pattern is validated against the connectivity constraint, which requires the active bridges to form a single connected region containing P1. Patterns that fail validation are rejected and regenerated. Shape families are used rather than uniform random switch states because uniform sampling does not produce usable patterns. To confirm this, 20,000 switch vectors were drawn at each of several values of an activation probability pp, with each of the 480 bits set independently. A draw was counted as valid if every active bridge lay inside a single connected region containing P1, which is the condition satisfied by every pattern in the dataset. At p=0.12p = 0.12, which activates 57.6 bridges on average and therefore matches the dataset mean of 58.6 closely, not one valid pattern appeared in 20,000 draws: the region reachable from P1 contained on average 1.5 patches, while almost all of the remaining switches lay in disconnected islands isolated from the feed port. The fragmentation persists as pp increases. Between p=0.02p = 0.02 and p=0.30p = 0.30 the mean number of active bridges rises from 9.6 to 144, yet the mean region reachable from P1 grows only from 1.1 to 4.1 patches, so adding metal does not add radiating structure. Connectivity is reached only once the lattice is close to saturated: at p=0.6p = 0.6 the mean region covers 180 of the 256 patches, which is an almost solid sheet rather than a shaped resonator, and its size can no longer be chosen. The pattern size is also fixed by uniform sampling, since the number of active bridges is binomial with standard deviation 480p(1−p)\sqrt{480p(1-p)}, which is about 7 switches at p=0.12p = 0.12, whereas the dataset must span three to several hundred switches to place resonances across the full band. The connected region therefore has to be constructed rather than sampled.

A pattern is described by two related quantities, both of which are used later in this work. The first is the number of active switches, meaning the horizontal and vertical bridges turned on between adjacent patches. This is the quantity the design variable controls directly, since the 480-bit vector is a list of switch states. The second is the number of patches joined into the P1-connected region, which is the metal that carries current and radiates. Turning on a switch either attaches a further patch to the region or closes a loop inside it, so the two counts rise together without being equal. Across the dataset the mean is 58.6 active switches with a range from 3 to 409, and the connected region ranges from 4 patches to 225, with half the dataset lying between 26 and 48 patches. The distribution of active switch counts is shown in Figure 3(c). The number of independent loops in a region, given by the number of switches minus the number of patches plus one, averages 21.7 and is used as one of the graph features described in Section IV.

Three histograms showing dip bandwidth, resonance frequency, and active switches

Figure 3. Dataset statistics. (a) Distribution of the −3-3 dB bandwidth of detected dips, with a median of 20 MHz. (b) Distribution of resonance frequency across all detected dips. (c) Distribution of the number of active switches per pattern.

Each generated pattern is simulated using the time-domain transient solver with open boundary conditions in CST Microwave Studio [30]. Copper and substrate are modelled as lossy materials with the parameters given in Section II. The frequency sweep covers 1.0 to 6.0 GHz with 1001 uniformly spaced samples, giving a resolution of 5 MHz. The reflection coefficient S11S_{11} at P1 is the observable metric, used throughout this work.

A total of 31,659 patterns were generated and simulated. Of these, 27,730 patterns, or 87.6%, produced an S11S_{11} curve with a minimum value below the −10 dB threshold, which is the conventional return loss level at which an antenna is considered matched. In a further step, only patterns with at least one local minimum below −10 dB that also has a prominence of at least 1 dB were retained, which distinguishes a true resonance from a curve that only marginally crosses the threshold. In this step 39 patterns were removed whose S11S_{11} touched −10 dB at isolated noisy points without forming a proper resonance, leaving a working dataset of 27,691 patterns containing 83,482 detected resonances in total.

The resonances in this dataset are narrowband. As shown in Figure 3(a), the median −3 dB bandwidth of an individual dip is 20 MHz, which is a small fraction of the 5000 MHz sweep. The narrowest resonances sit at 10 MHz, which is the resolution limit of the frequency sweep, and the widest reaches 230 MHz. Each pattern has on average 3.01 deep dips. The deepest dip of a pattern has a mean depth of −21.8 dB, while the mean taken across all detected dips is −17.3 dB. Due to this narrowbandness property the classification approach is utilised, since a resonance that occupies so little of the frequency axis is easily missed by a model trained to reproduce the full curve, described in Section IV. The distribution of resonance frequencies is shown in Figure 3(b). The 1.0 to 1.5 GHz interval contains 3,240 dips and the 5.5 to 6.0 GHz interval contains 12,406, so resonances are present across the full band with higher density at the upper end, which is consistent with the geometric tendency of small connected regions to resonate at higher frequencies. The dataset is partitioned into training, validation, and test subsets in the ratio 80:10:10 using a pseudo-random permutation with a fixed seed of 42.

This gives 22,152 training patterns, 2,769 validation patterns, and 2,770 test patterns. The same partition is used for all subsequent experiments. The summary of the dataset is shown in Table 1.

QuantityValue
Patterns simulated in CST31,659
Patterns after S11<−10S_{11} < -10 dB filter27,730
Patterns after prominence filter27,691
Training subset22,152
Validation subset2,769
Test subset2,770
Mean active switches per pattern58.6
Range of active switches3 to 409
Mean patches connected to P138
Range of connected patches4 to 225
Mean deep dips per pattern3.01
Total deep dips detected83,482
Mean depth of deepest dip per pattern−21.8-21.8 dB
Mean depth across all dips−17.3-17.3 dB
Median dip −3-3 dB bandwidth20 MHz
Range of dip −3-3 dB bandwidth10 to 230 MHz
Frequency-sweep range1.0 to 6.0 GHz
Frequency-sweep samples1001

Table 1. Dataset summary.

Forward Surrogate Model

Regression of the full reflection-coefficient curve was examined first, on patterns produced by the same generator and simulation protocol. A convolutional model trained to reproduce ∣S11∣|S_{11}| at all 1001 frequency points reached a mean absolute error of 1.2 dB, which on its own suggests an accurate model. Its predicted resonance, however, lay on average 1.46 GHz from the simulated one, nearly a third of the 5 GHz band. The low error was earned on the flat baseline, which covers most of the sweep, while the resonance, the feature most relevant to inverse design, was not located.

This motivates the central change made in this work. The forward problem is reformulated from regression of the curve to multi-label classification of resonance location. Instead of reproducing ∣S11∣\lvert S_{11}\rvert at every frequency point, the model predicts, for each of 100 frequency bins, the probability that a deep resonance is present. The training signal is thereby concentrated on the resonances rather than spread over the baseline. This allows the model to focus on the feature that matters for inverse design and provide the inverse search with a direct estimate of where a given pattern resonates.

Soft Label Construction

The forward model is trained to predict, from a switch pattern, whether a deep resonance is present in each of 100 frequency bins. The 1.0 to 6.0 GHz band is uniformly partitioned into 100 bins, so each bin is 50 MHz wide. The label vector for a pattern is built from its CST-simulated S11S_{11} curve by detecting local minima.

A local minimum at frequency fdf_d is treated as a deep resonance if its depth is below −10-10 dB and its prominence is at least 1 dB. The prominence is the vertical distance from the minimum up to the lower of the two adjacent local maxima, so a shallow dip lying on the shoulder of a deeper one is correctly assigned a small prominence and rejected. The −3-3 dB bandwidth BWd\mathrm{BW}_d is the width of the frequency interval over which the curve remains within 3 dB of the value at the minimum. A Gaussian bump is added to the label vector for each detected resonance, with standard deviation

σd=max⁡(BWd2,σmin⁡),(1)\sigma_d = \max\left(\frac{\mathrm{BW}_d}{2}, \sigma_{\min}\right), \tag*{(1)}

where σmin⁡=40\sigma_{\min} = 40 MHz. The floor σmin⁡\sigma_{\min} is the operative term for almost every resonance in this dataset. The bandwidth term exceeds it only when BWd>80\mathrm{BW}_d > 80 MHz, which holds for 0.5% of detected dips, so 99.5% of labels are Gaussians of fixed 40 MHz width. This is deliberate. With a median dip bandwidth of 20 MHz and a bin width of 50 MHz, a label scaled directly to the resonance bandwidth would be narrower than a single bin and would place nearly all of its mass on that one bin, leaving no gradient signal in the neighbouring bins.

Through this floor, every resonance, however narrow, is guaranteed to be spread across several adjacent bins, while the bandwidth term widens the labels of the small minority of broad resonances. When a pattern has more than one detected resonance, the Gaussian contributions are combined by element-wise maximum, so the nearest resonance determines the label at each bin. The label at bin bb, whose centre frequency is fbf_b, is

yb=max⁡dexp⁡(−(fb−fd)22σd2),(2)y_b = \max_d \exp\left(-\frac{(f_b-f_d)^2}{2\sigma_d^2}\right), \tag*{(2)}

where the maximum is taken over all detected resonances dd in the pattern and σd\sigma_d is given by (1). The bin nearest a true resonance therefore has yby_b close to unity, neighbouring bins receive partial credit, and bins far from any detected resonance have yby_b close to zero. A label attains unity exactly when a resonance coincides with a bin centre, which occurs for about one dip in ten because the bin centres and the frequency sweep share the same 5 MHz grid. The construction is illustrated in Figure 4.

Soft-label construction, shown on an illustrative response

Figure 4. Soft-label construction, shown on an illustrative response. (a) Two resonances are detected, one of 20 MHz and one of 140 MHz bandwidth. (b) Each resonance contributes a Gaussian whose width follows (1); the narrow resonance takes the 40 MHz floor while the broad one takes half its bandwidth, and where the two overlap the elementwise maximum of (2) is retained. (c) The combined field sampled at the 100 bin centres gives the label vector used for training. The curve in (a) is drawn for illustration; the bin width, the σ\sigma floor, and the combination rule are those used throughout.

Input Representation

The forward model has two inputs. The first is an image tensor of shape (3,16,16)(3, 16, 16). The horizontal-bridge state is held in channel 0, the vertical-bridge state in channel 1, and a binary mask marking the patches connected to P1 in channel 2. The port-connectivity information is thereby made directly available to the convolutional layers, rather than having to be derived from the bridge channels alone.

The second input is a vector of five scalar features describing the P1-connected subgraph: the number of patches in the component, the number of bridges inside the component, the longest hop distance from P1 measured in graph edges, the number of independent cycles in the component computed as edges minus nodes plus one, and the area of the bounding box enclosing the connected patches. The connectivity mask and the five features are obtained from a single breadth-first traversal of the bridge graph starting at P1, in time linear in the number of active bridges. The features are standardized using the mean and standard deviation of the training subset.

Network Architecture

The forward model is a convolutional network with an auxiliary scalar branch. The architecture is shown in Figure 5 and its hyperparameters are listed in Table 2.

Forward surrogate network architecture

Figure 5. Forward surrogate network architecture. The image branch processes the (3,16,16)(3, 16, 16) input through a stem and three dilated residual blocks. Three pooling operations and a graph-feature layer produce four descriptors that are concatenated and passed through a three-layer head to yield 100 frequency-bin logits.

ComponentValue
Image input shape(3,16,16)(3, 16, 16)
Graph-feature input5 scalars
Stem output channels96
Residual blocks3 (dilations 1, 2, 4)
Convolutions per block2
1×11 \times1 projection96 to 192 channels
PoolingAttention + GAP + GMP
Graph-feature layer5 to 32
Concatenated feature dim608
Head hidden sizes384, 256
Output dim100
Dropout (head)0.25
ActivationGELU
NormalizationBatchNorm2d
Trainable parameters879,397
OptimizerAdamW
Initial learning rate10−310^{-3}
Learning-rate scheduleCosine to 10−510^{-5}
Weight decay10−510^{-5}
Minibatch size128
Maximum epochs150
Early-stopping patience15 epochs
LossWeighted BCE, (3)

Table 2. Forward model hyperparameters.

The image branch begins with a stem convolution, in which the 3 input channels are expanded to 96 feature channels. The stem is followed by three residual blocks [31] with dilation rates 1, 2, and 4, each block containing two dilated convolutions with batch normalization [32] and the GELU activation [33]. The receptive field is expanded by dilated convolutions [34] without downsampling the feature map, so the full spatial resolution is preserved throughout. With a kernel size of 3, the cumulative receptive field is 3×33 \times3 after the stem, 7×77 \times7 after the first block, 15×1515 \times15 after the second, and 31×3131 \times31 after the third.

The second block therefore already spans almost the entire lattice and the third covers it with margin, thereby providing a hierarchical representation of local, mid-range, and full-board structures whose spatial extent is only 16 cells. The feature map is then projected to 192 channels by a 1×11 \times1 convolution. The 192-channel feature map is summarized using three parallel pooling operations. In learnable attention pooling [35], a soft mask is computed over the 16×1616 \times16 spatial positions, followed by a weighted sum of the feature vectors. Global average pooling computes the mean across spatial positions, while global maximum pooling extracts the maximum activation for each channel. These operations provide complementary representations: attention pooling produces a weighted spatial representation, average pooling captures the mean activation level, and maximum pooling retains the strongest response in each channel.

Each pooling operation produces a 192-dimensional descriptor. The five-feature scalar input is passed through a single linear layer with GELU activation that maps it to 32 features. The four descriptors, three pools of 192 each together with the 32 graph features, are concatenated into a 608-dimensional vector and passed through a head of three linear layers with hidden sizes 384 and 256 and an output of 100 logits. Dropout is applied after each of the two hidden layers. The network has 879,397 trainable parameters.

Training Procedure

The output of the model is a vector of 100 logits, one per frequency bin. A sigmoid is applied to each logit to obtain the predicted probability that a deep resonance is present in that bin. The model is trained with a per-bin weighted binary cross-entropy loss.

Let nn index the training sample within a minibatch of size NN, let bb index the frequency bin out of B=100B = 100, let zn,bz_{n,b} be the model logit, let yn,by_{n,b} be the soft label defined by (2), let σ(⋅)\sigma(\cdot) denote the logistic sigmoid, and let wbw_b be a per-bin weight. The loss is

L=−1NB∑n∑b[wbyn,blog⁡σ(zn,b)+(1−yn,b)log⁡(1−σ(zn,b))](3)\mathcal{L} = -\frac{1}{NB}\sum_n\sum_b \left[w_b y_{n,b}\log\sigma(z_{n,b}) + (1-y_{n,b})\log(1-\sigma(z_{n,b}))\right] \tag*{(3)}

The first term inside the sum is the cost incurred for assigning low probability to a bin that contains a resonance. The second term is the cost incurred for assigning high probability to a bin that does not. The positive-class term is scaled by the factor wbw_b to compensate for class imbalance. Low-frequency bins have far fewer positive examples than high-frequency bins, so without compensation the loss would be dominated by the more abundant high-frequency positives.

The per-bin weight is computed from the marginal positive-class frequency of the training subset as

wb=1−pbpb,pb=1∣Dtrain∣∑n∈DtrainI[yn,b>0.5],(4)w_b = \frac{1-p_b}{p_b}, \qquad p_b = \frac{1}{\lvert D_{\mathrm{train}}\rvert}\sum_{n\in D_{\mathrm{train}}}\mathbb{I}[y_{n,b}>0.5], \tag*{(4)}

where pbp_b is the fraction of training patterns whose label at bin bb exceeds 0.5 and I[⋅]\mathbb{I}[\cdot] is the indicator function. The form of wbw_b in (4) is the standard positive-class weighting used with binary cross-entropy under class imbalance: the expected gradient contribution from positive and negative examples is equalised in each bin. A large wbw_b is assigned to bins with pbp_b close to zero, while bins with pbp_b close to 0.5 are assigned wbw_b close to unity. Before the weight is formed, pbp_b is clamped to the interval [10−4,1−10−4][10^{-4}, 1-10^{-4}], which bounds wbw_b at 10410^{4} and avoids a division by zero in any bin that contains no positive example in the training subset.

The network is trained with the AdamW optimizer [36], a variant of Adam [37] with decoupled weight decay, at an initial learning rate of 10−310^{-3}, with weight decay of 10−510^{-5}, and minibatch size 128. The learning rate is annealed by a cosine schedule [38] to a floor of 10−510^{-5} over a maximum of 150 epochs. Mixed-precision training is used. Early stopping is applied on the validation mean area under the receiver operating characteristic curve with a patience of 15 epochs. The best validation checkpoint, reached at epoch 53, is retained for all subsequent evaluations.

Inverse Design Methodology

Search Space and Validity Constraint

The inverse-design task is a constrained discrete optimization over the 480-bit switch space. A pattern is valid when its active bridges form a single connected region that contains P1 and holds at least two patches. The reason for the constraint is that the response is measured at P1, so any metal not joined to P1 carries no fed current and does not contribute to the reflection coefficient. Every pattern in the training set has this property, as reported in § 2.1, so restricting the search in the same way also keeps the candidates inside the distribution the forward model was trained on.

The constraint is maintained throughout the search rather than applied as a filter afterwards. Moves that add a bridge preserve connectivity by construction, and moves that remove a bridge are checked before they are accepted, so every state the chain visits is a valid pattern.

Objective Function

Let s∈{0,1}480s \in\{0,1\}^{480} denote a switch state, let ftf_t be the requested target frequency, and let btb_t be the index of the 50 MHz bin that contains ftf_t. Let P(s;b)∈[0,1]P(s;b) \in[0,1] be the probability that bin bb contains a deep resonance, as predicted by the trained forward model of Section IV after applying the sigmoid to the corresponding logit. Let n(s)n(s) be the number of patches in the P1-connected region of ss. The objective maximized by the inverse search is

J(s)=max⁡b∈{bt−1,bt,bt+1}P(s;b)−λ⋅n(s)256.(5)J(s) = \max_{b \in\{b_t-1,b_t,b_t+1\}} P(s;b) - \lambda\cdot\frac{n(s)}{256}. \tag*{(5)}

The first term is a localised score. The highest predicted resonance probability within a three-bin window of ±50\pm50 MHz around the target is returned. The window absorbs small frequency offsets between the search and the target, which would otherwise penalize a candidate whose resonance falls just outside the centre bin.

Large connected regions are penalized by the second term. The denominator 256 is the total number of patches on the board, so n(s)/256n(s)/256 lies between 0 and 1. The penalty coefficient is λ=0.25\lambda= 0.25, which reduces the score by about 10−310^{-3} for each additional patch. That is small compared with the score range of [0,1][0,1] but large enough to separate candidates of otherwise similar score.

A hard cap of nmax⁡=80n_{\max} = 80 patches is also applied while moves are generated. The cap is set by training coverage rather than by fabrication. The training subset averages 38 connected patches per pattern, and only 0.7% of the dataset exceeds 150 patches, so a search allowed to drift well above that range would rely on extrapolated predictions from the forward model. The cap and the penalty together keep the search inside the region of input space where the model has support.

Simulated Annealing

The search uses a simulated annealing chain [39] that operates directly on the switch state. Each move changes one bridge. Three kinds of move are used.

  • A grow move activates a bridge between a patch already in the region and an adjacent patch that is not, which adds one patch to the region.

  • A close-loop move activates a bridge between two patches that are both already in the region, which adds a cycle but no patch.

  • A remove move deactivates one active bridge chosen at random. The resulting region is then tested, and the move is kept only if the region is still connected to P1 and still holds at least two patches. Up to six bridges are tried in turn, and if none of them can be removed the step is skipped.

At each step the kinds of move that have at least one candidate bridge available are collected, and one kind is drawn uniformly from that set. Grow moves are excluded once the region has reached nmax⁡n_{\max} patches, so the region never exceeds the cap.

The proposed state is accepted or rejected by the Metropolis rule. Let JcurJ_{\mathrm{cur}} be the objective value of the current state, JnewJ_{\mathrm{new}} that of the proposed state, and TT the current temperature. The acceptance probability is

pacc=min⁡(1,exp⁡(Jnew−JcurT)).(6)p_{\mathrm{acc}} = \min\left(1,\exp\left(\frac{J_{\mathrm{new}}-J_{\mathrm{cur}}}{T}\right)\right). \tag*{(6)}

Improving moves are always accepted. Worsening moves are accepted with a probability that falls as the gap Jnew−JcurJ_{\mathrm{new}}-J_{\mathrm{cur}} grows and as TT drops. Early in the chain, the relatively high TT increases the probability of accepting worsening moves, allowing transitions away from local maxima. As the TT decreases, the acceptance probability for worsening moves declines, and the search increasingly favours objective-improving transitions. The temperature follows a geometric schedule from T0=1.0T_0 = 1.0 to Tf=0.01T_f = 0.01 across the SS steps of the chain, and is reduced at every step whether or not a move was made. For each target frequency, M=10M = 10 chains are run independently from different random seeds, each for S=5000S = 5000 steps. Each chain is initialised with a random P1-rooted region of 4 to 20 patches. The best state seen during the run is returned by each chain rather than its final state, so a poor move late in the chain cannot spoil the result. The seed of chain mm at target ftf_t is 44+31m+⌊10ft⌋44 + 31m + \lfloor10f_t \rfloor, which gives a distinct seed for every chain and target. Algorithm 1 summarises the procedure.

Candidate Ordering

The candidate states returned for a given target frequency are often variations on a single topology. They are therefore Algorithm 1 Simulated annealing for inverse design.

Require: target frequency ftf_t, chains MM, steps SS, temperatures T0T_0 and TfT_f, patch cap nmax⁡n_{\max}, forward model probabilities P(⋅;⋅)P(\cdot;\cdot).

Ensure: MM candidate states.

  1. for m=1m = 1 to MM do

  2. Initialise ss as a random P1-rooted region of 4 to 20 patches.

  3. sbest←ss_{\mathrm{best}} \leftarrow s

  4. for t=1t = 1 to SS do

  5. T←T0⋅(Tf/T0)t/ST \leftarrow T_0 \cdot(T_f/T_0)^{t/S}

  6. Collect the available move kinds at ss, excluding grow if n(s)=nmax⁡n(s) = n_{\max}.

  7. Draw one kind uniformly and apply a move of that kind to obtain s′s'.

  8. if no valid move was found then

  9. continue

  10. end if

  11. Δ←J(s′)−J(s)\Delta\leftarrow J(s') - J(s) from (5)

  12. if Δ>0\Delta> 0 or rand⁡()<exp⁡(Δ/T)\operatorname{rand}() < \exp(\Delta/T) then

  13. s←s′s \leftarrow s'

  14. if J(s)>J(sbest)J(s) > J(s_{\mathrm{best}}) then

  15. sbest←ss_{\mathrm{best}} \leftarrow s

  16. end if

  17. end if

  18. end for

  19. cm←sbestc_m \leftarrow s_{\mathrm{best}}

  20. end for

  21. Order {c1,…,cM}\{c_1,\ldots,c_M\} by greedy max-min Hamming distance.

  22. return the ordered candidates.

ordered by greedy max-min selection before being passed to full-wave simulation. The first candidate is the one with the highest objective value. Each following candidate is the one whose smallest Hamming distance to the candidates already chosen is largest. The most distinct topologies are placed first by the ordering, so that a run that simulates fewer candidates than the chains produced still covers several regions of the pattern space. In the results reported here the number of candidates simulated per target equals the number of chains, so all ten outputs were retained.

Results

Forward model performance

The forward model is evaluated on the held-out test set. Two metrics are reported. The first is the area under the receiver operating characteristic curve (AUC), computed for each frequency bin. For a fixed bin bb, AUC is the probability that a randomly chosen pattern with a real resonance in bin bb receives a higher predicted score than a randomly chosen pattern without one. A binary target is obtained by thresholding the soft label at 0.5. AUC does not depend on a decision threshold. The mean per-bin AUC is 0.737, against a chance value of 0.5, and no bin falls below chance.

The second metric is precision at KK (P@K). For a fixed target bin, P@K is the fraction of the KK patterns with the highest predicted score at that bin that actually contain a resonance in the bin. This matches the way the model is used inside the inverse search, where candidates are ranked by their predicted score at the target bin and only the top-ranked ones are taken forward. P@K is averaged over 20 evenly spaced target bins with centre frequencies from 1.125 to 5.875 GHz. The values of P@1, P@5, P@10, and P@20 are 0.550, 0.470, 0.405, and 0.393. Precision falls as KK grows, as expected when a ranked list is extended.

Pipeline verification by full-wave simulation

The complete pipeline is verified by full-wave simulation of 220 generated candidates. Twenty-two target frequencies between 1.2 and 5.8 GHz are used, with 10 candidates generated per target by the inverse search of § 4. All 220 candidates are simulated in CST under the same protocol as the training dataset described in § 2.1. Two quantities are measured for each candidate: where its resonance falls relative to the requested frequency, and how deep that resonance is. Measuring the first requires the resonance to be located, and a detection threshold of −8-8 dB is used for that purpose. This is deliberately looser than the −10-10 dB matching criterion applied to the training dataset in § 2.1. A candidate whose S11S_{11} dips to −9-9 dB at exactly the requested frequency has placed its resonance correctly, and discarding it would hide a success of the search rather than reveal a failure. The −10-10 dB criterion is retained as the matching criterion and is reported separately as the depth statistic below, so that frequency placement and match quality are assessed independently. Furthermore, each of the frequency accuracy figures quoted below is changed by no more than two percentage points when the analysis is restricted to the candidates that reach −10-10 dB.

For each simulated curve, a search window of ±1\pm1 GHz is placed around the candidate’s target frequency, and the deepest local minimum below −8-8 dB inside that window is recorded as the produced resonance. A candidate is counted as successful when such a resonance is found. Of the 220 candidates, 166 are successful, a rate of 75%.

Frequency accuracy is reported over these 166 candidates. Figure 6(a) shows the requested target frequency against the simulated dip frequency for each one, with the dashed line marking a perfect match, the shaded bands marking the ±100\pm100 MHz and ±300\pm300 MHz tolerance regions, and the marker colour giving the depth of the resonance. Of these candidates, 38% produce a resonance within 100 MHz of the requested target, 66% within 300 MHz, and 86% within 500 MHz, with a median error of 152 MHz. Figure 6(b) gives the cumulative distribution of the frequency error, read as the fraction of candidates whose dip falls within a given distance of the request. These figures treat every candidate as a separate result.

Inverse design results: scatter plot, cumulative distributions, generated switch patterns, and simulated reflection coefficients

Figure 6. Inverse design results. (a) Requested against simulated dip frequency for the 166 successful candidates. (b) and (c) Cumulative distributions of the frequency error and the dip depth. (d) and (e) Two generated switch patterns. (f) and (g) Their simulated reflection coefficients.

The method is not used in that way: ten candidates are returned for each request, all ten are simulated, and the one whose simulated dip lies closest to the request is taken as the answer. On that basis the median error across the 22 requests is 23 MHz, as reported in § 5-C. The choice is made on the simulated response rather than on the surrogate prediction, so no candidate is set aside on the strength of a model output.

The depths of the 166 resonances range from −8.0-8.0 to −38.2-38.2 dB, and their cumulative distribution is given in Figure 6(c). Of these candidates, 80% reach a depth of −10-10 dB or deeper and 42% reach −15-15 dB or deeper, with a median of −14-14 dB. Using the −10-10 dB matching criterion alone, 132 of the 220 simulated candidates, or 60%, produce a matched resonance within the search window. The colour scale in Figure 6(a) shows that these deeper resonances occur at target frequencies across the whole band rather than being confined to one part of it.

Two generated designs are shown in Figure 6(d) and (e), for requested frequencies of 3.8 and 3.0 GHz. Active patches are drawn in blue, closed switches as solid segments, and open switches with a break at the centre. Figure 6(f) and (g) give the corresponding simulated reflection coefficients. In each, the shaded bands mark ±50\pm50 MHz and ±100\pm100 MHz about the simulated dip. The solid vertical line marks the center dip of simulated S11S_{11}, and the dashed vertical line marks the resonance frequency predicted by the surrogate.

The solver is replaced by the surrogate inside the search, so the key issue is the resulting performance loss relative to a search that used the solver directly. To answer it, the annealing chain of Section V was rerun with the scoring function of (5) replaced by a full-wave simulation at every step. The move set, the connectivity constraint, the patch cap, and the temperature schedule are unchanged; only the quantity being maximised differs. Each candidate is simulated in CST and scored by the depth of the deepest point within ±50\pm50 MHz of the target, so both searches reward a resonance near the target, one through its simulated depth and the other through its predicted probability.

A matched comparison at the search length used here is not possible. In the surrogate-guided search 50,000 states are evaluated per request, and simulating each one would take several months of solver time for a single target. The solver-guided search was therefore given 200 simulations per target, which at an average of about 3.5 minutes per simulation, including data export, is about 11.5 hours per request and is at the upper end of what is practical in a design workflow. The same 22 target frequencies were used, giving 4,400 simulations. The study required 253 hours (∼\sim10.5 days) of continuous solver time.

The cost of the proposed method separates into two parts. The search evaluates 50,000 states through the surrogate and completes in 2.2 minutes, with no solver calls. Verifying the ten candidates it returns takes 10 full-wave simulations, or about 35 minutes, giving roughly half an hour per request.

All timings were measured on the same workstation, with an NVIDIA RTX 3090 used for both the full-wave transient solver and the surrogate. Building the training dataset required 31,659 simulations, a cost incurred only once. The solver-guided search needs 200 simulations for every request, so about 160 requests would cost as many simulations as the whole dataset.

Figure 7(a) plots the median frequency error of both methods against the number of simulations spent, with the corresponding solver time on the upper axis. The proposed method reaches a median error of 23 MHz after its 10 simulations. The solver-guided search does not reach that level within its 200 simulations, ending at 35 MHz, so it uses twenty times the solver budget and roughly twenty times the wall-clock time without matching the result. For the first 50 simulations it has typically found no resonance inside the target window at all, because the objective is flat wherever no dip is present and the chain has no gradient to follow. A graded probability is returned by the surrogate at every state, so the search it drives always has a direction.

Comparison of median frequency error and best result for each target frequency

Figure 7. Comparison with a solver-guided search over the same 22 targets. (a) Median frequency error against the number of full-wave simulations spent per request, with the equivalent solver time on the upper axis. Shaded bands are the interquartile range. Where no resonance has yet been found inside the search window the error is plotted at the 1 GHz window limit. (b) Best result for each target frequency.

The two methods are compared target by target in Figure 7(b). The proposed method is closer at 15 of the 22 targets, with a median error of 23 MHz against 35 MHz and a mean of 32 MHz against 115 MHz. Every request is answered within 300 MHz, while the solver-guided search fails at two: at 2.4 GHz a structure resonating at 1.48 GHz is returned and at 2.6 GHz one resonating at 3.60 GHz, both well matched but at the wrong frequency.

The solver-guided search produces substantially deeper resonances, with a median depth of −30.7 dB compared with −13.6 dB for the proposed method. This difference is expected from the objectives used in the two searches. The objective in (5) maximises the predicted probability of a resonance occurring within the target bin, but does not directly optimise resonance depth. A shallow resonance at the correct frequency can therefore achieve a high objective value. The solver-guided search, in contrast, directly maximises the resonance depth and consequently favours deeper resonances.

Discussion

The forward surrogate reaches a mean per-bin AUC of 0.737, while the pipeline it drives answers all 22 requests within 300 MHz through the best of their ten candidates. A moderate AUC is sufficient here because the surrogate is used only inside the search, never as a stand-alone predictor. States are scored inside a search that is already restricted to valid patterns, so only candidates that differ by one bridge have to be ordered, and the response of an arbitrary pattern never has to be predicted. The chain is held inside the range of pattern sizes present in the training set by the size penalty, so the states that are scored stay close to the data the surrogate was fitted on. The final candidate performance is assessed through full-wave simulation, which is why the method is reported against 220 simulated candidates rather than against the surrogate’s own test performance.

An extension of the method to two-port filter operation on the same board is reported in the remainder of this section, followed by the limitations of the present work.

Extension to two-port filter operation

A second mode of operation is supported by the same board. A second port, P2, sits at the mirror of P1 on the opposite edge of the lattice, and is not used in the antenna work above. If the active bridges are made to join both ports rather than only P1, the structure becomes a two-port network and the observable of interest changes from the reflection coefficient alone to the pair S11S_{11} and S21S_{21}. This was investigated as an extension of the present work, with the aim of building a surrogate that could later be inverted to return a structure for a requested passband. The attempt is reported here because its outcome bears on the argument made in Section I about which kinds of response this board and this modelling approach can support.

Patterns were produced by a generator built on the same principles as the antenna generator of Section III. Every pattern begins from a conducting path that joins P1 to P2, either running straight along the port row or meandering within one row of it, so a two-port connection exists by construction and no simulation is spent on structures that cannot conduct. Additional metal is then placed around that path from a fixed set of shapes: open and shorted stubs of bounded length, square and concentric rings of bounded size, and vertical pillars held clear of the main line so that they couple to it without shorting into it. The board is therefore not filled at random. Each pattern is a two-port structure with a transmission path and a limited number of loading elements. A total of 24,535 such patterns were simulated in CST under the same protocol as the antenna dataset, with S11S_{11} and S21S_{21} recorded from 1.0 to 6.0 GHz.

What the Dataset Contains

A filter must meet three conditions at once. It must pass the requested band with low insertion loss, it must be matched over that band so the power enters the structure rather than reflecting, and it must reject outside the band. Each pattern was therefore tested for a contiguous region in which S11S_{11} falls below −10-10 dB and S21S_{21} stays above −3-3 dB at the same frequencies, and then for rejection below −15-15 dB in a 500 MHz window placed 100 MHz beyond each edge of that region.

The two conditions give very different counts. A passband satisfying the first condition is present in 17,610 of the 24,535 patterns, or 71.8%. When the rejection condition is added, 386 patterns are left, or 1.6%. Passbands are therefore common on this board, whereas rejection beside them is rare. The rejection test is a local one, since it examines only 500 MHz on either side of the passband rather than the whole sweep, and the board still fails it in 98% of cases.

What the Surrogate Learned

A forward surrogate was trained on the full set of 24,535 patterns to predict the S11S_{11} and S21S_{21} curves directly, as a regression over the 1001 frequency points. Unlike the antenna task, no reformulation was needed. The filter response varies gradually across the band rather than concentrating in a few narrow dips, so a regression objective is not dominated by a flat baseline and the difficulty described in § 1 does not arise in the same form.

A mean absolute error of 1.85 dB on S11S_{11} and 4.02 dB on S21S_{21} is reached by the trained model over a held-out test set. Two patterns and their responses are shown in Figure 8, with the prediction plotted on the same axes. The main features of both curves are placed at the right frequencies, including the transmission nulls and the edges of the passband, but the predicted S21S_{21} level differs from the simulation by up to 10 dB in places.

Two-port filter operation on the same board

Figure 8. Two-port filter operation on the same board. (a) and (b) Two patterns with both ports connected. (c) and (d) The corresponding simulated S11S_{11} and S21S_{21}, with the surrogate prediction on the same axes and the shaded band marking the detected passband.

The S21S_{21} error is too large for the surrogate to drive an inverse search. A passband requires S21S_{21} above −3-3 dB and a stopband requires it below −15-15 dB, so the two are separated by 12 dB. A mean error of 4 dB is a third of that separation. Transmission can therefore be reported as high or low in a given region, but a passband cannot be reliably told apart from a stopband, which is the minimum requirement for a filter surrogate.

Interpretation

Two limits are visible, and both are present in the data before any question of network architecture arises. The first is coverage. Only 1.6% of the dataset contains a passband with adjacent rejection, so the target response is rarely represented in the training data. The second is resolution. The broad form of S11S_{11} and S21S_{21} is learned by the surrogate, but not to the accuracy that separating a passband from a stopband requires.

The two modes impose different requirements. A resonance is a local feature. It occupies a small part of the frequency axis, and a design method only has to place that one feature correctly, which is what the classification formulation of Section IV does and what the full-wave verification of § 5 confirms. A filter is a specification on the response across the whole sweep, requiring transmission in one region and rejection in another at the same time. On this board that outcome is rare, and it is not reached by the same modelling approach.

LIMITATIONS

The proposed work with the antenna method has four limitations. First, every candidate is anchored at P1 by the pattern generator of § 2.1, which produces a training distribution biased toward the left half of the lattice. Configurations that resonate below 2 GHz often need to extend further across the grid, which the generator covers poorly. A generator that grows from arbitrary seed locations would broaden the region of the board the surrogate has seen and would likely improve performance at the low end of the band.

Second, the bridges are modelled as ideal copper segments. In a real reconfigurable board, non-ideal switches such as PIN diodes or radio-frequency micro-electromechanical systems are used, which add parasitic resistance, capacitance, and inductance. These parasitics, together with the DC bias lines needed to control each switch, shift the resonances and widen the bandwidth by amounts that depend on the switch technology. Because the surrogate learns from simulated data, these effects can be included by simulating the training patterns with a lumped switch model, with no change to the classification formulation or the search. A fabricated prototype with measured scattering parameters is required to establish the gap between idealised simulation and hardware.

Third, the inverse search optimises for resonance frequency only. Radiation pattern, polarization, and gain are not part of the objective and would require additional surrogate outputs or a selection step applied to the candidate set. Fourth, the presence of a resonance at the requested frequency is scored by the objective of (5), but not its depth. The consequence is shown by the comparison in § 5 where deeper resonances are produced by a solver-guided search optimising depth directly.

Conclusion

In this paper, an inverse design method for a 16×1616 \times16 reconfigurable pixel antenna with 480 binary switches has been presented and verified by full-wave simulation. The method has two parts. The forward task is posed as multi-label classification over a discretized frequency axis rather than as regression of the reflection-coefficient curve, which suits a response whose useful content is a small number of narrow dips in a wide sweep. The inverse task is a simulated annealing search over the switch state in which the connectivity constraint is maintained at every step, so the chain only ever visits patterns that are valid antennas, and a size penalty keeps it inside the region of the input space the surrogate was trained on.

The method was tested end to end by generating candidates for 22 target frequencies between 1.2 and 5.8 GHz and simulating all 220 of them in CST. All 22 requests were answered within 300 MHz by the best of their ten candidates, and 80% of the detected resonances reached below −10 dB. A mean per-bin AUC of 0.737 on a held-out test set is reached by the forward surrogate that guides the search, which is sufficient to rank the valid candidate patterns evaluated during the search and identify configurations whose resonant frequencies approach the requested target. A search that scores every candidate with the full-wave solver instead of the surrogate reaches a median error of 35 MHz after 200 simulations per request, against 23 MHz after 10 for the method presented here.

The same board was also examined in two-port operation, where the objective is a filter response rather than a resonance. That extension did not reach the same standard. Passbands are common in the simulated data, appearing in 71.8% of patterns, but only 1.6% also reject the frequencies immediately beside the passband, and a surrogate trained on the full set predicts S21S_{21} with a mean absolute error of 4.02 dB, which is too coarse to separate a passband from a stopband. Placing one narrow feature at a requested frequency is a task this board and this method perform well, while specifying the response across the whole sweep is not.

Three directions follow from this work. The first concerns coverage of the board. Every training pattern is anchored at P1 by the pattern generator, so extending it to grow from arbitrary seed locations would broaden the region of the board the surrogate has seen and should improve performance at the low end of the band, where large connected regions are needed. The second concerns hardware. The switches are modelled as ideal copper, so a fabricated prototype with measured scattering parameters is required to establish the gap between simulation and hardware. Such a board would also allow the training data to be collected by measurement, since each switch state can be set electronically and its reflection coefficient recorded with a vector network analyser in seconds. A surrogate trained on measured data would then include the effects of real switches and bias lines directly. The third is filter operation. A much larger dataset weighted toward topologies that produce rejection would be needed, together with a training objective that targets the passband and stopband levels directly rather than the mean error across the curve, before the same approach could be applied.

Nadeem Rather

Figure 9.

NADEEM RATHER is a Research Engineer with the Wireless Communications and AI Research Group at Tyndall National Institute, University College Cork (UCC), Ireland. He received the Ph.D. degree in electronic engineering from UCC in 2024. His research focuses on artificial intelligence (AI)-enabled wireless systems, chipless radio-frequency identification (RFID) sensing, electromagnetic system optimisation, and advanced antenna technologies. His research interests also include deep learning for radar-based sensing, reconfigurable intelligent surfaces, and intelligent electromagnetic structures for next-generation Internet of Things applications. He has received several research distinctions, including the Wrixon Research Excellence Bursary in 2024, the Best Application of AI in a Student Project award at the AI Awards Ireland 2024, and the Best Student Paper Award at the International Workshop on Antenna Technology. He also received the Best Technology Demonstration Award at the VistaMilk Internal Conference in 2023 and was recognised as a runner-up in Tyndall National Institute’s Postgraduate Publication of the Year competition in 2023.

Holger Claussen

Figure 10.

HOLGER CLAUSSEN Prof. Holger Claussen is Head of the Wireless Communications Laboratory at Tyndall National Institute, and Professor of Wireless Communications at University College Cork and Trinity College Dublin. At Tyndall, he is building up research teams in RF, Access, Protocols, AI, and Quantum Systems to invent the future of Wireless Communications. Previously, he led the Wireless Communications Research Department of Nokia Bell Labs in Ireland and the US until 2020. He received the 2014 World Technology Award in Communications Technologies for innovative work of “the greatest likely long-term significance” for his contributions to Small Cell Networks. Holger received his Ph.D. degree from the University of Edinburgh, UK in 2004. He is author of the book “Small Cell Networks”, more than 140 journal and conference publications, 87 granted patent families, and 39 filed patent applications pending. He is Member of the Royal Irish Academy, Fellow of the IEEE, and Fellow of Trinity College Dublin.

Lester Ho

Figure 11.

LESTER HO is a Principal Scientist at the Tyndall National Institute in Dublin where he leads the Future Access team. Previously, he spent 20 years performing research in wireless networks in Nokia Bell Labs, where he was a Distinguished Member of Technical Staff. There, he performed research in the areas of self-optimizing networks, cellular architectures, and network energy efficiency. He has invented the self-optimization techniques and scalable architectures that enabled the commercial deployment of small cells, where many of his innovations are used in live cellular networks today. His current research interests are in network optimization at the physical to network layers, O-RAN, machine learning, quantum repeater networks, and wireless network simulations.

References

  1. [1]C. G. Christodoulou, Y. Tawk, S. A. Lane, and S. R. Erwin, “Reconfigurable antennas for wireless and space applications,” Proceedings of the IEEE, vol. 100, no. 7, pp. 2250–2261, 2012.DOI
  2. [2]J. Costantino, Y. Tawk, S. E. Barbin, and C. G. Christodoulou, “Reconfigurable antennas: Design and applications,” Proceedings of the IEEE, vol. 103, no. 3, pp. 424–437, 2015.
  3. [3]S. V. Hum and J. Perrisseau-Carrier, “Reconfigurable reflectarrays and array lenses for dynamic antenna beam control: A review,” IEEE Transactions on Antennas and Propagation, vol. 62, no. 1, pp. 183–198, 2014.DOI
  4. [4]R. L. Haupt, “An introduction to genetic algorithms for electromagnetics,” IEEE Antennas and Propagation Magazine, vol. 37, no. 2, pp. 7–15, 1995.DOI
  5. [5]L. Jing, M. Li, and R. Murch, “Compact pattern reconfigurable pixel antenna with diagonal pixel connections,” IEEE Transactions on Antennas and Propagation, vol. 70, no. 10, pp. 8951–8961, 2022.DOI
  6. [6]M. Di Renzo, A. Zappone, M. Debbah, M.-S. Alouini, C. Yuen, J. de Rosny, and S. Tretyakov, “Smart radio environments empowered by reconfigurable intelligent surfaces: How it works, state of research, and the road ahead,” IEEE Journal on Selected Areas in Communications, vol. 38, no. 11, pp. 2450–2525, 2020.
  7. [7]C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis, and I. Akyildiz, “A new wireless communication paradigm through software-controlled metasurfaces,” IEEE Communications Magazine, vol. 56, no. 9, pp. 162–169, 2018.arxiv.org/abs/1806.01792
  8. [8]M. P. Bendsøe and N. Kikuchi, “Generating optimal topologies in structural design using a homogenization method,” Computer Methods in Applied Mechanics and Engineering, vol. 71, no. 2, pp. 197–224, 1988.
  9. [9]R. E. Christiansen and O. Sigmund, “Inverse design in photonics by topology optimization: Tutorial,” Journal of the Optical Society of America B, vol. 38, no. 2, pp. 496–509, 2021.DOI
  10. [10]S. Molesky, Z. Lin, A. Y. Piggott, W. Jin, J. Vučković, and A. W. Rodriguez, “Inverse design in nanophotonics,” Nature Photonics, vol. 12, no. 11, pp. 659–670, 2018.
  11. [11]K. Deb, A. Pratap, S. Agrawal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.DOI
  12. [12]P. R. Wiecha, A. Arbouet, C. Girard, and O. L. Muskens, “Deep learning in nano-photonics: Inverse design and beyond,” Photonics Research, vol. 9, no. 5, pp. B182–B200, 2021.DOI
  13. [15]H. M. El Miliani and T. Naous, “Machine learning in antenna design: A comprehensive overview,” IEEE Access, vol. 11, pp. 103 890–103 915, 2023.
  14. [16]J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B. G. DeLacy, J. D. Joannopoulos, M. Tegmark, and M. Soljačić, “Nanophotonic particle simulation and inverse design using artificial neural networks,” Science Advances, vol. 4, no. 6, p. eaar4206, 2018.
  15. [17]D. Liu, Y. Tan, E. Khoram, and Z. Yu, “Training deep neural networks for the inverse design of nanophotonic structures,” ACS Photonics, vol. 5, no. 4, pp. 1365–1369, 2018.
  16. [18]S. Sang and R. D. Murch, “An efficient approach for optimizing frequency reconfigurable pixel antennas using genetic algorithms,” IEEE Trans. Antennas Propag., vol. 62, no. 2, pp. 609–620, 2014.
  17. [19]S. Shen, Y. Sun, S. Song, D. P. Palomar, and R. D. Murch, “Successive Boolean optimization of planar pixel antennas,” IEEE Trans. Antennas Propag., vol. 65, no. 2, pp. 920–925, 2017.DOI
  18. [20]T. Wu, B. Liu, Q. Hua, M. O. Akinsolu, D. P. Buch, J. J. Adams, M. Imran, P. Lazaridis, R. Pei, and P. Excell, “An efficient method for complex digitally coded antenna design based on evolutionary computation and machine learning techniques,” IEEE Trans. Antennas Propag., 2025.DOI
  19. [21]A. Gupta, E. A. Karahan, C. Bhat, K. Sengupta, and U. K. Khankhoje, “Tandem neural network based design of multiband antennas,” IEEE Transactions on Antennas and Propagation, vol. 71, no. 8, pp. 6308–6317, 2023.DOI
  20. [22]L.-Y. Xiao, W. Shao, F.-L. Jin, B.-Z. Wang, and Q. H. Liu, “Inverse artificial neural network for multiobjective antenna design,” IEEE Transactions on Antennas and Propagation, vol. 69, no. 10, pp. 6651–6659, 2021.DOI
  21. [23]S. Chen, G.-H. Sun, and K. Wang, “Inverse design of microstrip antennas based on deep learning,” Electronics, vol. 14, no. 13, p. 2510, 2025.DOI
  22. [24]N. Rather, H. Claussen, and L. Ho, “Machine learning assisted inverse design of pixelated mmWave patch antennas,” in Proc. IEEE Int. Symp. Personal, Indoor and Mobile Radio Communications (PIMRC), 2026, pp. 1–6.arxiv.org/abs/2608.23469
  23. [25]J. P. Jacobs, “Accurate modeling by convolutional neural-network regression of resonant frequencies of dual-band pixelated microstrip antenna,” IEEE Antennas and Wireless Propagation Letters, vol. 20, no. 12, pp. 2417–2421, 2021.DOI
  24. [26]N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. A. Hamprecht, Y. Bengio, and A. Courville, “On the spectral bias of neural networks,” in Proceedings of the 36th International Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 97, 2019, pp. 5301–5310.arxiv.org/abs/1806.08734
  25. [27]R. P. Ribeiro and N. Moniz, “Imbalanced regression and extreme value prediction,” Machine Learning, vol. 109, no. 9, pp. 1803–1835, 2020.DOI
  26. [28]D. M. Pozar, Microwave Engineering, 4th ed. Hoboken, NJ, USA: Wiley, 2011.
  27. [29]C. A. Balanis, Antenna Theory: Analysis and Design, 4th ed. Hoboken, NJ, USA: Wiley, 2016.
  28. [30]Dassault Systèmes, “CST Studio Suite,” Software, 2023, sIMULIA, Dassault Systèmes, Vélizy-Villacoublay, France.
  29. [31]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778.arxiv.org/abs/1512.03385
  30. [32]S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International Conference on Machine Learning (ICML), 2015, pp. 448–456.
  31. [33]D. Hendrycks and K. Gimpel, “Gaussian error linear units (GELUs),” arXiv preprint arXiv:1606.08415, 2016.
  32. [34]F. Yu and V. Koltun, “Multi-scale context aggregation by dilated convolutions,” in International Conference on Learning Representations (ICLR), 2016.arxiv.org/abs/1511.07122
  33. [35]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems (NeurIPS), 2017, pp. 5998–6008.arxiv.org/abs/1706.03762
  34. [36]I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations (ICLR), 2019.
  35. [37]D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in International Conference on Learning Representations (ICLR), 2015.
  36. [38]I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” in International Conference on Learning Representations (ICLR), 2017.arxiv.org/abs/1608.03983
  37. [39]S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated annealing,” Science, vol. 220, no. 4598, pp. 671–680, 1983.

Paper details

vlnc:W_6e26dk96
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents