Uday Kiran Khankhoje

In this paper, we introduce novel machine learning based techniques to design multi-band microstrip antennas as per user specifications over a broad range of frequencies. The approach involves the design and training of a neural network for approximating the electromagnetic simulations of antennas, the so-called 'forward' problem. Here, the antenna is parameterized in terms of a checker-board pattern of metallic sub-patches. Additionally, a second 'tandem' neural network is also designed, which takes the user specification of a desired return-loss spectrum and returns an antenna structure. We explore the various machine learning innovations that are required in order for this approach to succeed. Our approach makes way for rapid designs of multi-band antennas, which is otherwise known to be a tedious task requiring vast domain knowledge.

Detecting element failure in a phased array antenna plays a crucial role in ensuring a communication system's efficiency since faulty elements lead to a degradation of the array performance. Under the assumption that only a few antenna elements are faulty, fault diagnosis can be accomplished by applying compressive sensing techniques to solve the resulting system of equations. We present a fault diagnosis method of a linear antenna array, where the measurements are taken at a fixed location, and excitations are optimized. We solve the compressive sensing problem that leads to a reduction in the number of measurements required for successful diagnosis using the optimized excitations. We show how the excitations can be optimized for fault detection in the presence of cosine squared field pattern of an antenna element in a linear array.

We review the recent demonstration of multi-point simultaneous sensing based on External Modulation Brillouin Correlation Domain Analysis (EM-BOCDA), and the use of gradient descent algorithm for accurate estimation of the Brillouin Frequency Shift (BFS).

Quantitative microwave imaging using inverse scattering is a promising technique for biomedical imaging. Given the ill-posed nature of the inverse problem, the use of efficient regularization techniques is essential in order to come up with meaningful solutions to the imaging problem. In this paper, we propose a novel regularization technique that is based on an iterative divide-and-unite algorithm of the imaged domain. Multi-scaling procedures have been proposed earlier, where the object domain is iteratively divided based on heuristic criteria. We take a different route, where, starting from a single coarse pixel, the domain is divided into finer pixels based on heterogeneity in the gradient of the cost function. An inexpensive algebraic reconstruction technique is then applied to estimate the values of the finer pixels. Subsequently, an unite step is performed to combine pixels with similar values of dielectric contrast. The power of this method is that it keeps the number of reconstructed pixels at a minimum and allows for nonlocal pixels, as also seen in level-set based reconstructions. Implementation of the algorithm shows a significant reduction in converging time, the cost function and the total shape error.

In this paper, we introduce inverse design of nearly arbitrary planar antenna structures with a deep convolutional neural network (CNN) modeling that allows rapid and accurate prediction of antenna performance (scattering parameters and radiation patterns). Quite distinct from prior efforts of ML-based antennas with fixed template geometries and finite degrees of freedom, this approach of generalizing to arbitrary planar structures opens up a new design space for antenna structures with properties beyond what can be achieved with antennas optimized from a finite library. By eliminating complex time consuming electromagnetic simulations with an ML-based approach, we propose an inverse design with evolutionary algorithms that allows a much larger search space than classical genetic algorithm based approaches. We demonstrate this methodology with simulation and measurement results of inverse designed compact, broadband and multi-band planar antennas operating at RF (2-5 GHz) and mmWave (20-40 GHz).

We propose an irrotationality-preserving total variation algorithm to solve the two dimensional (2D) phase unwrapping problem, which occurs in Interferometric Synthetic Aperture Radar (InSAR) imaging and other problems. Total variation methods aim at denoising the phase derivatives to reconstruct the absolute phase. We supplement these methods by adding an additional constraint driving the curl of the gradient of the 2D phase map to zero, i.e. imposing the irrotationality of the gradient map by suitably constructing a cost function which we then minimize. We test our method and compare with existing methods on several synthetic surfaces specific to the problem of InSAR imaging for different noise levels. We report better estimates of unwrapped phase maps for the terrains simulated and for all noise levels with a two-fold improvement in terms of root mean square (RMS) error in high noise level scenarios.

Antenna fault diagnosis for phased antenna arrays is an important research area since faulty elements deteriorate the expected field pattern, leading to degraded performance in various applications. While several compressive sensing-based techniques have been proposed, they rely on a simplified array factor formula, ignoring mutual coupling effects among antennas. We show that this assumption can lead to poor diagnosis in the presence of significant mutual coupling by using two popular models - the average embedded element pattern and a port-level coupling matrix approach. Also, we optimize the antenna excitations to minimize the mutual coherence of the system measurement matrix, leading to a reduced number of measurements required for fault diagnosis. Our simulation results indicate that accounting for the effect of mutual coupling results in a far more reliable diagnosis. In addition, our framework is executed using a single measurement probe fixed in space, thus making a step toward practical fault diagnosis techniques that can be deployed on antenna array systems.

In this paper, we propose an efficient method to estimate the electromagnetic field distribution as a function of space over the region of interest from the knowledge of field at a few points. For this purpose, we use the Huygens' principle, where the field at any point in the region of interest is expressed in terms of the tangential field on the surface of the scattering object. The tangential fields are expressed as a Fourier series expansions on the scatterer boundaries; these Fourier coefficients are estimated by taking measurements at random spatial points around and outside the scatterer. Once the Fourier coefficients across the boundary of the scatterer are known, we evaluate the field at arbitrary points using the Huygens' principle. Truth data is generated at all locations using either the BI method or a Finite Difference Frequency Domain method. The error is computed between the predicted field (based on noisy measurements) and truth data on a grid of square points around the scattering object. We study this error as a function of noise level, the number of basis functions, and the number of measurements after performing suitable Monte Carlo averages over several realizations of measurement points.

We review the recent demonstration of multi-point simultaneous sensing based on External Modulation Brillouin Correlation Domain Analysis (EM-BOCDA), and the use of gradient descent algorithm for accurate estimation of the Brillouin Frequency Shift (BFS).

Estimation of the accurate value of Brillouin frequency shift (BFS) in a Brillouin optical correlation domain analysis (BOCDA) is challenging due to the contributions to the Brillouin gain spectrum (BGS) from the locations other than corresponding to the correlation between the pump and the probe. In this scenario, we demonstrate optimal post-processing algorithms to retrieve the BFS accurately. We first demonstrate a linear approximation based approach to estimate the BFS with an accuracy of <1 MHz. This approach needs to be modified for situations where simultaneous sensing is carried out from two or more locations, or with lock-in detection methods. A gradient descent method is proposed and demonstrated in such cases where the component corresponding to the second harmonic of the modulation frequency is used to optimally recover BFS. The method is tested for its performance at different locations of correlation peaks and for perturbation frequency range of >500 MHz, under different SNR conditions. The error in estimation is found to be less than 1.7 MHz across the entire frequency range for an SNR as low as 5 dB. The algorithm is also validated by comparison with experimental data. The proposed algorithm effectively increases the range and sensitivity of measurement for BFS estimation even when multiple locations are monitored simultaneously.