Can CNNs be Used to Predict Multi-Physics Simulations of Microwave Induced Damage to Basalt?

Understanding the fundamentals of microwave induced damage by conducting multi-physics simulations is a complex task that requires the integration of several physical phenomena and models. In such scenarios, convolutional neural networks (CNNs) can be used to extract intrinsic behavior and patterns from large complex datasets generated from simulations and experiments. The goal of this research is to use temperature images obtained from COMSOL simulations, density, relative permittivity, and power of microwave to train and validate CNN. The data obtained from COMSOL simulations is pre-processed to make it suitable for use as input to the CNN. This process involves normalizing and resizing the image dataset. The data is split into two sets, namely the training and validation set. The CNN is trained using the training set where the CNN learns the inherent characteristics of the data. Upon successful completion of the training, the CNN is tested using previously seen and unseen data. The CNN perceptions indicate a high prediction accuracy:
$76\% < \epsilon^{\prime} < 95\%$
and
$82\% < \epsilon^{\prime\prime} < 96\%$
for image sizes (64, 64) and (256, 256) respectively. The use of CNN for understanding microwave induced damage can provide a powerful mechanism of gaining insight into the complex phenomena and identify novel ways of manipulating and controlling the parameters involved in a microwave induced damage.