Machine learning approach for identification and tracking of coherent structures in turbulent fluids and plasmas

Abstract

In a fusion reactor, coherent structures of hot and dense plasma can drift radially outwards due to the conditions of the edge plasma and can cause erosion of the outer walls. This erosion can release impurities into the plasma and harm equipment at the walls. This thesis presents two methods of tracking blobs in the boundary region of fusion experiments. The first model is a simple Long Short-Term Memory model with few layers. The second model is a more advanced transformer structure, with more depth and parameters. The performance of the models is evaluated with synthetic data, and compared on experimental data with a pre-trained model. The generation of synthetic data with different distributions in amplitude, size, velocity, and the number of blobs is also presented to better understand when the model is viable. Calculations of the velocities, amplitudes, and sizes of structures found in synthetic and experimental data are presented, where results on experimental data are compared to published results from earlier studies. The transformer-based model shows promising results on synthetic data that has low intermittency. It is shown that higher parameter variation results in worse model performance. Predictions on experimental data show that the model has some problems, including differentiating between blobs and predicting large structures from 0-values. Size and velocity estimates in experimental data are found to be in the same order of magnitude as in previous studies. The Long Short-Term Memory model shows promising results in segmenting the shape of the blobs, but lacks the capacity to differentiate them correctly.