20 February 2025 in IEEE Journal on Multiscale and Multiphysics Computational Techniques
Antenna synthesis is becoming increasingly challenging with tight requirements for C-SWAP (cost, size, weight and power) reduction while maintaining stringent electromagnetic performance specifications. While machine learning approaches are increasingly being explored for antenna synthesis, they are still not capable of handling large shape sets with diverse responses. We propose a branched deep convolutional neural network architecture that can serve as a drop-in replacement for a full-wave simulator (it can predict the full spectral response of reflection co-efficient, input impedance and radiation pattern). We show the utility of such models in surrogate-assisted evolutionary optimization for antenna synthesis with arbitrary specification of targeted response. Specifically, we consider the large shape set defined by the set of 16-vertexes polygonal patch antennas and consider antenna synthesis by specifying independent constraints on return loss, radiation pattern and gain. In contrast to online surrogates, our approach is an offline surrogate that is objective-agnostic; trained once, it can be used over multiple optimizations whereby the model training costs become amortized across multiple synthesis requests. Our approach outperforms evolutionary optimizations relying on full-wave solver-based fitness estimation. Specifically, we report the design, fabrication and experimental characterization of three polygon -shaped patch antennas, each fulfilling different objectives (narrow band, dual-band & wide-band). The reported methodology enables rapid synthesis (in seconds), produces verifiable sound designs and is promising for furthering data-driven design methodologies for electromagnetic wave device synthesis.
5 May 2025 in Optics Express
The diverse range of shapes enabled by modern nanofabrication techniques makes it challenging to identify the most optimal design for a desired optical response. While deep learning-based approaches are being increasingly explored for inverse design (especially for optical metasurfaces), they have mostly been limited to small design subsets that constrain the shapes, thicknesses, or parameters like pitch, incident angle, etc. Scalability of such techniques to the full design space accessible to modern nanofabrication remains a challenge due to the difficulty of training models which retain acceptable generalization across wider design spaces. We explore the possibility of using active learning techniques for sample-efficient model training and surrogate-driven inverse design. Specifically, we consider the inverse design of periodic optical metasurfaces with 36 diverse shape classes and broad variations in thickness and pitch. Our results demonstrate that an active learning-driven approach to inverse design can give comparable performance to random-dataset trained models with a substantial (up to 82%) reduction in training dataset generation. The inverse design capability is seen across diverse spectral filter optimization tasks with the added benefit of providing multiple solutions in a single run. As a model- and training-paradigm-agnostic dataset reduction technique, our approach enables deep learning-based inverse design to scale to increasingly larger design spaces.
13 August 2025 in IEEE Antennas and Wireless Propagation Letters
We propose a novel, scalable deep Bayesian optimization (BO) methodology for designing antennas with a large number of design degrees of freedom. Conventional BO approaches in antenna design have relied on Gaussian process (GP) surrogates, which limits its scalability to higher-dimensional design spaces. To address this limitation, we propose (1) a deep neural network (DNN) surrogate with Monte Carlo (MC) dropout for efficient multi-output Bayesian inference, (2) an active learning strategy to construct an informative initial dataset, and (3) a hybrid expected improvement–differential evolution (EI-DE) acquisition scheme balancing global exploration with local exploitation for efficient sample selection. Applied to a 52-variable ultra-wideband (UWB) antenna design scenario, the proposed method achieves a 75% reduction in computational cost compared to a conventional DE algorithm while achieving similar solution quality. It also outperforms some existing surrogateassisted optimizers, reducing computation time by over 40% while yielding superior designs. The proposed deep learning driven BO framework offers a promising direction for antenna synthesis.
11 April 2025 in TechRxiv
We propose a novel scalable deep Bayesian optimization (BO) methodology for designing antennas with a large number of design degrees of freedom. Conventional BO approaches in antenna design has relied on Gaussian process (GP) surrogates, which limits its scalability to higher-dimensional design spaces. To address this limitation, we propose (1) a deep neural network (DNN) surrogate with Monte Carlo (MC) dropout for efficient multi-output Bayesian inference, (2) an active learning strategy to construct an informative initial dataset, and (3) a hybrid expected improvement-differential evolution (EI-DE) acquisition scheme balancing global exploration with local exploitation for efficient sample selection. Applied to a 52-variable ultra-wideband (UWB) antenna design scenario, the proposed method achieves a 76% reduction in computational cost compared to a conventional DE algorithm while achieving similar solution quality. It also outperforms some existing surrogateassisted optimizers, reducing computation time by over 42% while yielding superior designs. The proposed deep learningdriven BO framework offers a promising direction for antenna synthesis.