Automatic estimation of heat loads distribution on STRIKE through multi-layer perceptrons

To characterize the negative ion beam accelerated from the Source for Production of negative Ion of Deuterium Extracted from Radio Frequency plasma (SPIDER), the estimation of the heat flux on its diagnostic calorimeter STRIKE is needed. The calorimeter consists of 16 tiles made of carbon fiber-carbon composite (CFC). Since the front observation of the calorimeter tiles (facing the ion source) is not possible, the reconstruction of the impinging beam distribution is mandatory. The task is an inverse problem, where the temperature data on the opposite side with respect to the ion source are used to reconstruct the heat flux coming from the ion source, on the front side. Traditional methods, when applied to experimental data, require iterative approaches and are time-consuming, especially when dealing with the 1280 beamlets received by fully operative STRIKE. To overcome this drawback, Neural Network-based models are proposed to reconstruct the heat flux by estimating the parameters representing its distribution. In this work, Multi-Layer Perceptron (MLP) models are trained, validated and tested on a dataset of 163 experimental infrared (IR) images. The assumption of a gaussian distribution of the temperature patterns is adopted, and two different parametrizations of the IR images are evaluated. The heat flux estimated by the MLPs, when used to reconstruct the temperature on the back side of the tile, provides an average pixel error below 6% with respect to the experimentally measured temperature data.