Link to paper
The full paper is available here.
You can also find the paper on PapersWithCode here.
Abstract
- Finding antenna designs that meet frequency requirements and are optimal is a critical component in designing next generation hardware.
- The process is non-trivial because the objective function is nonlinear and sensitive to design changes.
- EM simulations are slow and expensive with commercial software.
- CZP is a sample-efficient and accurate surrogate model to estimate scattering coefficients in the frequency domain of a given 2D planar antenna design.
- CZP uses a novel image-based representation for antenna topology and attention-based neural network architectures.
- CZP outperforms baselines in terms of test loss and can find 2D antenna designs with only 40k training samples.
Paper Content
Introduction
- Next generation of Metaverse computing devices offer exciting possibilities for immersion in the digital world
- Antenna design is challenging due to physical volume constraints and demand for broader spectrum coverage
- Devices contain up to 10-20 antennas in a small form factor
- WiFi 6E extends the WiFi high band to the entire 6GHz spectrum
- Computational antenna design relies on full-wave electromagnetic simulations
- Process is expensive in terms of simulation and engineer time
- Two approaches to antenna surrogate modeling: approximate physics-driven simulation or data-driven methods
- Leverage machine learning to efficiently learn a surrogate model
- Antenna search space is high dimensional and relationship between antenna topology and resonance profile is highly non-linear
- Proposed novel image-based representation of antennas geometries
- Resonance characteristic of an antenna can be written analytically as the ratio of two complex-valued polynomials
- Proposed transformer architecture to tokenize the image representation and predict the complex valued zeros and poles of the S scattering matrix
Antenna preliminaries
- Defined key notions and concepts
- Introduced antenna design problem
- Presented derivation based on EM solvers
- Theoretical innovation of CZP architecture
Antenna parameterization
- We consider 2D planar antennas to demonstrate an approach
- An antenna design consists of a substrate, ground plane, discrete port, and front-side metallic patches
- Combining all these specifications, we have an overall design choice vector which determines the antenna’s frequency response
- Antenna engineers aim to find the right design choice so that specific antenna design targets are met
- We consider an example antenna with an FR-4 substrate and 5 front-side metallic patches
- The relationship between an antenna’s topology and its responses is governed by Maxwell’s equations
- We can solve Maxwell’s equations by discretizing the space and converting the PDEs into ODEs
- We can compute the scattering coefficients S 11 (ω) by its definition
- We can also compute S 11 (ω) by directly performing a temporal Fourier transform for ODEs
- S 11 (ω) has an analytic form as the quotient of two complex polynomials with respect to frequency ω
- We can train a neural network to predict the zeros and poles, as well as a global constant c 0 from material properties h
Network architecture
- Proposed model includes a novel image representation for a 2D antenna
- Image representation inspired by mesh representations used by EM simulators
- Image-based transformer architecture predicts zeros and poles of scattering coefficients
Image representation
- Mesh-based finite element methods are used in many simulation tools.
- Mesh representations use the fact that an antenna’s resonance characteristics are related to its topology.
- Images can be used to learn a surrogate model as they contain spatial information.
- Adaptive meshing enables the simulation of systems unsolvable by traditional discretization methods.
- A three channel image representation is proposed to explicitly represent the boundaries and corners of substrate.
Surrogate model
- Proposed architecture for surrogate model to predict zeros and poles from image representation
- Augment image with two additional channels of linearly spaced x and y coordinates
- CNN takes augmented image as input and generates feature maps
- Filter-based tokenizer generates visual tokens to capture semantics such as relative boundary and corner locations
- Visual tokens passed through multi-layer transformer encoder
- Fully connected layer and non-linearity used to predict constant, zeros and poles
Model training
- Training CZP models requires no direct supervision of c 0 (h), zeros z(h), and poles p(h).
- Estimated S 11 (ω) is computed with c 0 , z, and p to match the ground truth.
- Model is trained via back-propagation to minimize Mean Squared Error (MSE).
- Frequency range is [0.2 − 7.0]GHz at increments of 0.1 (69 dimensions).
- Shrinkage loss variant of MSE is used to reduce error on crucial parts of the scattering coefficients.
Experiments
- Proposed image representation is better than coordinate-based inputs and binary image input.
- CZP formulation outperforms raw prediction with same transformer architecture.
- Transformer architecture outperforms CNN and MLP.
- CZP generalizes well to unseen antenna designs.
Surrogate modeling
- 48K samples used, 90% for training, 10% for testing, 10% of training set for validation
- Experiments demonstrate effectiveness of novel image representation and proposed CZP when using transformer architecture
- Coordinate-based method concatenates normalized bottom-left x, y coordinate of each patch with onehot vector
- Tokens projected into 256 dimensional vector by 2-layer MLP with hidden layer of width 256
- CZP improves over raw prediction by minimum of 4.6% and maximum of 28.0%
- Increasing transformer depth from 2 to 8 layers improves raw prediction for image and coordinate representations by 29.5% and 35.9%
- CZP and image representation benefit least from increasing complexity of model, raw prediction and coordinate representation benefit most
- 8-layer transformer with image input is 40%+ improvement on baselines
Optimization
- Proposed model can be used by optimization procedure to find antenna configurations with specific resonance characteristics
- CZP is more robust than raw prediction because it is smooth
- Antenna design framed as reinforcement learning problem
- State and action of Markov Decision Process defined
- Reward given at final timestep
- Optimization done using Soft Actor Critic
- CZP outperforms raw prediction in terms of success rate
- CZP is more robust to less data
Conclusion
- Proposed a novel surrogate model architecture for antenna design optimization
- Derived a theoretical form for the S 11 scatter coefficients
- Proposed a novel image representation inspired by existing mesh-based simulation techniques
- Demonstrated experimentally that the proposed model and image representation are significant advances
- Proposed model had significantly higher utility in terms of success rate for optimization of antenna design than baselines
- Future work will involve solving more complicated 2D problems and generalizing the proposed model and image representation to 3D antenna
- Future work will involve exploring other tokenization schemes and applying to linear PDEs in general
- ODE in the form of Eqn. 3 is diagonalizable and the logarithm of modulus of the scattering coefficients log |S 11 (ω)| can be written as a ratio of two complex polynomials
- Patch p 1 determines the location of the discrete port
- Image representation outperforms coordinates and CZP outperforms raw prediction
- Ablations of architectural components of the proposed model and baselines
- After all patches have been placed, the coordinates are converted to the image representation and the surrogate model predicts the frequency response log |S 11 (ω)|
- Compute reward components for each resonance target
- Double-sided Fourier transform and symmetric signal extension
- Comparison of predicted frequency response with an 8 layer transformer and image input on two test set examples
- Attention maps learned by the spatial attention component of the proposed transformer architecture