Front-end adaptive electronic modeling with neural networks for radioastronomy

Direct and inverse modeling with neural networks of a complex front end electronics for the S-AAIR project are evaluated. While direct model outputs S-parameters based on 4 electronic inputs parameters, inverse modeling has to output 4 optimal electronic parameters with respect to 5 S-parameters and noise constraints inputs. The advantages to be expected from neural netwoks include execution speed without loss of accuracy, and the possibility of modeling a system without any prior hypothesis based on measurements only, or any other source of information. This is a crucial step for designing a fully adaptive front-end electronics, with a single electronic board capable of adapting to several different observational and instrumental constraints like signal over noise, bandpass, dynamic range, power consumption, and several other possible features. In this first study, simulated data are used to evaluate the performances of such machine learning methods. The direct model is shown to reach less than 0.1 dB RMS error with respect to the simulated data, which is sufficiently accurate to be compatible with measurements-based modeling without distorsion. The inverse model is constructed by training a network on a pre-computed optimization function. It is able to output optimal electronics parameters satisfying a set of performances constraints without errors compared to the original function. Those first tests confirm neural networks are valid tools for complex RF electronic modeling, even with modest computing resources running basic networks topologies, and thanks to the availability of powerful associated algorithmy. There is thus several perpectives for improving and extending those first results.