Design Space and Frequency Extrapolation: Using Neural Networks

With the tremendous growth of the semiconductor industry, compute power and memory have become cheap and accessible. One interesting outcome of this growth has been the adoption of machine learning (ML) in several fields traditionally dominated by physics and mathematics [1]-[9]. Solving electrically large systems by analyzing their electromagnetic (EM), thermal, and mechanical behavior can be a time-and memory-intensive process. But, as is well known today, such analyses become inevitable with 1) the increase in operating frequencies, 2) the scaling in system and device size, and 3) the hybrid nature of different components packaged in close proximity. As system complexity increases, design cycles become longer since each product iteration requires the multivariable analysis of EM structures. Contemporary examples of such complexity are millimeter-wave (mmwave) systems, where multiple chiplets and microwave components are integrated on a single substrate or package [10], [11].