Diagnostic Data Integration Using Deep Neural Networks for Real-Time Plasma Analysis

In the study of magnetically confined plasma, working conditions are often extreme, meaning that the diagnostic system is subject to high noise, intermittent behaviour and malfunctions. In this framework, artificial intelligence can be of great help. The main focus of this thesis is the application of deep neural-networks to the diagnostic soft x-ray 3-arrays (DSX3), as to reconstruct the plasma temperature profiles, giving thus an aid to the diagnostic system. Firstly, the data collected by the DSX3 camera are analysed, in order to understand and classify the various types of temperature profiles that can be produced in a reverse field pinch machine such as RFX-mod. The screening of the profiles is done separating them via their magnetic topology, in either multiple helicity (MH) or quasi-single The RFX-mod machine, located at Consorzio RFX, Padova, Italy is one of the few devices in the world where the study of magnetically confined plasma is made possible both in reversed field pinch and tokamak configuration. When working with high temperature plasma, the conditions of operation for the diagnostic system are not always ideal meaning that it is frequent to encounter high noise, malfunction, intermittent behaviour and sometimes even failure of the data acquisition systems. In order to address these problems, we turn to the artificial intelligence world. The main focus of this thesis project is the application of deep neural networks (DNN) to the soft x-ray diagnostic in RFX-mod, in order to reconstruct the plasma temperature profiles, thus giving an aid to the diagnostic system in case of unreliable/missing data. Such a study is of particular importance towards near future campaigns in this device. The first step of the project has been focused on the analysis of almost 70,000 temperature profiles and magnetic data for more than 300 pulses. The target of the study was understanding the different types of temperature profiles that can be produced in RFX-mod associated with the multiple helicity or quasi-single helicity states, and characterizing them considering various metrics, such as the maximum electron temperature, the temperature profile width and its slope. The second step of the project aimed at training a neural network based on variational auto-encoder using RFX-mod database. This is a particular type of auto-encoder, capable of efficiently decoupling the important quantities present in the database, making the work of understanding the output of the model more lightweight. Using this methodology, a neural network model has been identified and its capability in reconstructing missing data has been assessed. This turns out to be particularly useful as a means of supplying reliable data in the case of failure or misbehaviour of the diagnostic system, such as in the case of saturation or blockage of the viewing channels, which is a recurrent problem in the specific case of the soft-x ray diagnostic. The capability of machine learning in reconstructing missing temperature data has been tested considering multiple use-cases. The advantage and the limits within this approach have been documented here.