Vijay K. Devabhaktuni

Artificial neural networks for RF and microwave design - from theory to practice

Neural-network computational modules have recently gained recognition as an unconventional and useful tool for RF and microwave modeling and design. Neural networks can be trained to learn the behavior of passive/active components/circuits. A trained neural network can be used for high-level design, providing fast and accurate answers to the task it has learned. Neural networks are attractive alternatives to conventional methods such as numerical modeling methods, which could be computationally expensive, or analytical methods which could be difficult to obtain for new devices, or empirical modeling solutions whose range and accuracy may be limited. This tutorial describes fundamental concepts in this emerging area aimed at teaching RF/microwave engineers what neural networks are, why they are useful, when they can be used, and how to use them. Neural-network structures and their training methods are described from the RF/microwave designers perspective. Electromagnetics-based training for passive component models and physics-based training for active device models are illustrated. Circuit design and yield optimization using passive/active neural models are also presented. A multimedia slide presentation along with narrative audio clips is included in the electronic version of this paper. A hyperlink to the NeuroModeler demonstration software is provided to allow readers practice neural-network-based design concepts.

A robust algorithm for automatic development of neural-network models for microwave applications

In this paper, we propose a robust algorithm for automating the neural network based RF/Microwave model development process. The algorithm can build a neural model starting with zero amount of training/test data, and then proceeding with neural network training in a stage-wise manner. In each stage, the algorithm utilizes neural network error criteria to determine additional training/test samples required and their location in model input space. The algorithm dynamically generates these new data samples during training, by automatic driving of simulation tools, e.g., OSA90, Ansoft-HFSS. Initially, fewer hidden neurons are used, and the algorithm adjusts the neural network size whenever it detects under-learning. Our technique integrates all the sub-tasks involved in neural modeling, thereby facilitating a more efficient and automated model building process. It significantly reduces the intensive human effort demanded by the conventional step-by-step neural modeling approach. The algorithm is demonstrated through MESFET and Embedded Capacitor examples.

Advanced microwave modeling framework exploiting automatic model generation, knowledge neural networks, and space mapping

In this paper, we propose an efficient Knowledge based Automatic Model Generation (KAMG) technique, aimed at generating microwave neural models of highest possible accuracy using fewest accurate data. The technique is comprehensively derived to integrate three distinct powerful concepts, namely, automatic model generation, knowledge neural networks and space mapping. We utilize two types of data generators - fine data generators that are accurate and slow (e.g., CPU-intensive 3D-EM simulators); coarse data generators that are approximate and fast (e.g., inexpensive 2D-EM). Motivated by the space-mapping concept, the KAMG utilizes extensive approximate data but fewest accurate data to generate neural models that accurately match fine data. Our formulation exploits a variety of knowledge network architectures to facilitate reinforced neural network learning from both coarse and fine data. During neural model generation by KAMG both coarse and fine data generators are automatically driven using adaptive sampling. The proposed technique is demonstrated through examples of MOSFET, and embedded passives used in multi-layer PCBs.

A robust algorithm for automatic development of neural network models for microwave applications

In this paper, we propose a robust algorithm for automating the neural network based RF/Microwave model development process. The algorithm can build a neural model starting with zero amount of training/test data, and then proceeding with neural network training in a stage-wise manner. In each stage, the algorithm utilizes neural network error criteria to determine additional training/test samples required and their location in model input space. The algorithm dynamically generates these new data samples during training, by automatic driving of simulation tools, e.g., OSA90, Ansoft-HFSS. Initially, fewer hidden neurons are used, and the algorithm adjusts the neural network size whenever it detects under-learning. Our technique integrates all the sub-tasks involved in neural modeling, thereby facilitating a more efficient and automated model building process. It significantly reduces the intensive human effort demanded by the conventional step-by-step neural modeling approach. The algorithm is demonstrated through MESFET and Embedded Capacitor examples.

Advanced microwave modeling framework exploiting automatic model generation, knowledge neural networks and space mapping

In this paper, we propose an efficient Knowledge based Automatic Model Generation (KAMG) technique, aimed at generating microwave neural models of highest possible accuracy using fewest accurate data. The technique is comprehensively derived to integrate three distinct powerful concepts, namely, automatic model generation, knowledge neural networks and space mapping. We utilize two types of data generators - fine data generators that are accurate and slow (e.g., CPU-intensive 3D-EM simulators); coarse data generators that are approximate and fast (e.g., inexpensive 2D-EM). Motivated by the space-mapping concept, the KAMG utilizes extensive approximate data but fewest accurate data to generate neural models that accurately match fine data. Our formulation exploits a variety of knowledge network architectures to facilitate reinforced neural network learning from both coarse and fine data. During neural model generation by KAMG both coarse and fine data generators are automatically driven using adaptive sampling. The proposed technique is demonstrated through examples of MOSFET, and embedded passives used in multi-layer PCBs.

Neural-network approaches to electromagnetic-based modeling of passive components and their applications to high-frequency and high-speed nonlinear circuit optimization

In this paper, artificial neural-network approaches to electromagnetic (EM)-based modeling in both frequency and time domains and their applications to nonlinear circuit optimization are presented. Through accurate and fast EM-based neural models of passive components, we enable consideration of EM effects in high-frequency and high-speed computer-aided design, including components geometrical/physical parameters as optimization variables. Formulations for standard frequency-domain neural modeling approach, and recent time-domain neural modeling approach based on state-space concept, are described. A new EM-based time-domain neural modeling approach combining existing knowledge in the form of equivalent circuits (ECs), with state-space equations (SSEs) and neural networks (NNs), called the EC-SSE-NN, is proposed. The EC-SSE-NN models allow EM behaviors of passive components in the circuit to interact with nonlinear behaviors of active devices, and facilitate nonlinear circuit optimization in the time domain. An automatic mechanism for EM data generation, which can lead to efficient training of neural models for EM components, is presented. Demonstration examples including EM-based frequency-domain optimization of a three-stage amplifier, time-domain circuit optimization in a multilayer printed circuit board, including geometrical/physical-oriented neural models of power-plane effects, and EM-based optimization of a high-speed interconnect circuit with embedded passive terminations and nonlinear buffers in the time domain are presented.

Prediction of protein-coding regions in DNA sequences using a model-based approach

Prediction of the protein-coding regions (exons) is one of the central issues of DNA sequence analysis. Most of the existing computational methods exploit the period-3 property of the coding-regions to distinguish exons from noncoding regions (introns). However, the current Discrete Fourier Transform (DFT) based methods are inadequate in predicting short exons. In this paper, we present a model-based exon detection approach using statistically optimal null filter. The proposed method employs a model of the period-3 characteristic to maximize signal-to-noise ratio, and least-squares optimization criteria to rapidly detect the presence of exons in the input DNA sequence. Through examples, it is shown that the proposed method is highly effective as compared to the DFT technique, especially in identifying short exons and successive exons separated by short introns.

Adjustable Bandwidth Filter Design Based on Interdigital Capacitors

In this letter, we explore a general-purpose microstrip coupling model for computer aided design of new bandpass filters. The J-inverter topology of the model facilitates the study of coupling behaviours of different microstrip structures leading to a quick comparison of their coupling strengths. Based on such comparison, microstrip filters with interdigital capacitors and etched slots yielding relatively higher coupling coefficients are designed, fabricated and tested. The filters exhibit relatively wider bandwidths which are easily adjustable by way of changing the geometrical parameters of interdigital capacitors and slots.

A two-stage neural network based technique for protein secondary structure prediction

Protein secondary structure prediction is one of the most important research areas in bioinformatics. In this paper, we propose a two-stage protein secondary structure prediction technique, implemented using neural network models. The first neural network stage of the proposed technique associates the input protein sequence to a bin containing its corresponding homologues. The second stage predicts the secondary structure of the input sequence utilizing a neural prediction model specific to the bin obtained from stage one. The strategy of binning allows for simplified and accurate neural models. This technique is implemented on the RS126 dataset and its prediction accuracy is compared with that of the standard PHD approach.

A New Approach to Analysis and Modeling of Esophageal Manometry Data in Humans

In this paper, we propose a new approach to the analysis and modeling of esophageal manometry (EGM) data to assist the diagnosis of esophageal motility disorders in humans. The proposed approach combines three techniques, namely, wavelet decomposition (WD), nonlinear pulse detection technique (NPDT), and statistical pulse modeling. Specifically, WD is applied to the filtering of the EGM data, which is contaminated with electrocardiography (ECG) artifacts. A new NPDT is applied to the denoised data leading to identification and extraction of diagnostically important information, i.e., esophageal pulses from the respiration artifacts. Such information is used to generate a statistical model that can classify the EGM patterns. The proposed approach is computationally effortless, thus making it suitable for real-time application. Experimental results using measured EGM data of 20 patients, including ten abnormal cases is presented. Comparison of our results with those from existing techniques illustrates the advantages of the proposed approach in terms of accuracy and efficiency.