Mauro Tucci

The spread of high-performance personal computers, frequently equipped with powerful Graphic Processing Units (GPUs), has raised interest in a set of techniques that are able to extract models of electromagnetic phenomena (and devices) directly from available examples of desired behavior. Such approaches are collectively referred to as Machine Learning (ML). A typical representative ML approach is the so-called “Neural Network” (NN). Using such data-driven models allows the evaluation of the output in a much shorter time when a theoretical model is available, or allows the prediction of the behavior of the systems and devices when no theoretical model is available. With reference to a simple yet representative benchmark electromagnetic problem, some of the possibilities and pitfalls of the use of NNs for the interpretation of measurements (inverse problem) or to obtain required measurements (optimal design problem) are discussed. The investigated aspects include the choice of NN model, the generation of the dataset(s), and the selection of hyper-parameters (hidden layers, training paradigm). Finally, the capabilities in the handling of ill-posed problems are critically revised.