Evolutionary Exploration of an Ultrasound Propagation Predictor Neural Net

To find an optimal treatment plan for an focused ultrasound based
procedure, a multitude of computationally expensive simulations need to
be evaluated, often thousands of times. 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. The learned
optimizer showed an excellent performance on the test set, and is
capable of generalization well outside the training examples, including
to much larger computational domains, and more complex source and sound
speed distributions. The utilized UNet architecture, however, was marked
by the authors as the uncertain part of the design with a real
possibility to improve upon. To explore their design more deeply and to
confirm their theories, we made an attempt to improve the solver by use
of an evolutionary algorithm, challenging the importance of different
building blocks and authors original choices and decisions regarding the
architecture. Two experiments using Cartesian Genetic Programming had
been ran, first to try and optimize the original nets architecture, and a
second one, to study the effects of the employed multi-resolution
encoding on precision of the network. Our exploitative experiments
managed to find a network with approximately an order of magnitude
better RMSE for the predictor, using the same validation set as the
original solver. The exploration evolution process then showed the use
of 4 resolution layers as valid, and provided topics for further
research on the effects of the memory blocks and convolution kernel
sizes.