Welcome to Kuntal's website!
Biography
Kuntal has a Ph. D. in Electrical Engineering from Arizona State University. He received his B. Tech. degree, also in Electrical Engineering, from the Indian Institute of Technology, Madras (now Chennai). He has worked in the area of modeling and simulation of integrated circuit components at Motorola Semiconductor and Texas Instruments as an individual contributor and as a manager. Most recently he has been focussing on the application of machine learning techniques for efficient simulation of higher order effects in semiconductor devices and model parameter extraction. Kuntal has 11 US patents and has published over 20 refereed papers in various journals and conferences. He has served on the technical committees of IEEE's International Electron Devices Meeting (IEDM) and Bipolar/BiCMOS Circuits and Technology Meeting (BCTM).
Projects
High Fidelity Models Using Machine Learning: With continuous advances in technology, it is becoming important for SPICE models to be able to simulate complex higher order effects in transistors. Unfortunately, relying on traditional physics based approaches to meet these requirements is prohibitive from both cost and time perspectives. A more practical approach is to use machine learning methods to enhance existing physics based formulations. This allows the well established core of these models to be retained while adding on capabilites to simulate new device phenomena. One example of this application is the addition of a neural layer over a traditional BSIM4 based core for simulating impact ionization effects in medium and high voltage LDMOS transistors. Details are available here. (link currently inactive)
Machine Learning Based Optimization: Another useful application of machine learning is in the area of parameter optimization. The traditional way to do this is using non-linear optimizers. A forward model is executed iteratively while adjusting the parameter vector to minimize an objective function, typically the RMS error between the model and some reference data. For SPICE model parameter extraction, this approach is typically very time consuming and requires continuous human supervision. A more efficient method using machine learning has been developed. This method first constructs an "inverse" model which relates desired model outputs to model parameters. MIMO neural networks are very suitable for building these inverse models. Once such a model is trained to sufficient accuracy, extracting parameters becomes a quick one-shot process. A key element for the success of this approach is feature engineering. Details are available here. (link currently inactive)
Data Visualization: Omniviz aims to be a general all-purpose data visualization and analysis tool. Written in Python and leveraging the tkinter and matplotlib libraries, OmniViz organizes data in an intelligent human logical manner. A prototype web version, OmniWeb for analysis and vizualization of financial data has been developed. The tool will be enhanced incrementally.
For a look at my other projects, click on the tabs at the top of this page.