Machine learning-based prediction of heat load on Tokamak divertor target plates

<sec>The SOLPS-ITER edge plasma simulation code has become a primary tool for designing divertor physics and predicting target plate heat load in fusion research. However, SOLPS-ITER-based divertor design requires not only substantial computational time but also intensive hardware resources, which fundamentally limits its application in advancing divertor optimization, particularly in large-scale fusion reactor divertor design. In this work, the machine learning method is used for the first time to predict the plasma parameters of the divertor target plate for HL-3, which provides a theoretical basis for predicting the heat load of divertor in large fusion reactor in the future. Based on the simulation of the edge plasma code SOLPS-ITER, we first build a database of HL-3 edge plasma parameters, including the upstream inner/outer midplane region and divertor target region. Then, we use the machine learning method and combine with the database to develop an artificial neural network model. Finally, the artificial neural network is used to train a model by using the boundary plasma parameters of the HL-3 device, and the heat load of the divertor target plate is predicted by the given upstream plasma parameters.</sec><sec>This work can effectively shorten the time of simulating edge plasma by SOLPS-ITER code from weeks, months or even half a year to several milliseconds. In this work, a multi-layer perceptron (MLP) model is established with different input parameters to predict the electron temperature, density, and parallel heat fluxes of the inner and outer divertor target plates. It is found that reasonably increasing upstream plasma parameters as inputs to the model can not only improve the model’s generalization ability and the accuracy of prediction (both reaching over 90%), but also verify the correlation between upstream plasma parameters and divertor target physical quantities. In addition, a more stable ResMLP model is established on the basis of MLP. This work demonstrates the feasibility of using the neural networks to predict the heat load of the divertor target plate.</sec>