Evolutionary NAS for Topology of an Acoustic Propagation Predictor

To find an optimal treatment plan for a High Intensity Focused Ultrasound surgery a multitude of computationally expensive simulations need to be evaluated, often thousands of times to obtain a precise treatment plan. Recent renaissance of machine learning technologies could provide a solution to this problem, as a recently published article presented a Physics Informed Neural Net to predict Acoustic Propagation through a human skull. While the net utilizes a UNet topology a is reasonably small, a multiple redundant parts are present within the design and the whole approach was to prove this approach is feasible. To validate this net for use in HIFU treatment plan optimization loop, an attempt was made to try and find a different architecture for the net, minimizing the number of parameters while preserving the precision with use of a combination of genetic algorithm and cartesian genetic programming.