ON THE USE OF MACHINE LEARNING TO MAINTAIN THE SPATIAL GRID QUALITY IN SOLVING GAS DYNAMICS PROBLEMS
A. V. Babanov, A. V. Voevodin, A. N. Shcherbakov
VANT. Ser.: Mat. Mod. Fiz. Proc 2022. Вып.2. С. 53-60.
The problem of automatically correcting a difference grid while numerically solving gas dynamics problems is considered. In the Eulerian-Lagrangian code MIMOZA, a spatial grid has an unsatisfactory quality because of distorted difference grid’s lines having the Lagrangian flag which prevents from the reconstruction of nodes on these lines upon completion of the Lagrangian stage in the numerical solution of gas dynamic equations. Such feature of nodes is used, as a rule, to select contact boundaries of materials coinciding with lines of the difference grid. If the difference grid deformation is strong, there occurs the time point during the problem solution, when it is required to cancel Lagrangian flag of nodes. This operation in the MIMOZA code is manually performed, as usual, and introduces to the numerical solution the factor of uncertainty of the Lagrangian flag cancellation time. The paper authors resolve the problem of identifying defective nodes of the difference grid having Lagrangian flag and timely selecting another type of the grid reconstruction by using the machine learning technology. A template input dataset is presented for training the neural network, which characterizes premises for the generation of defective configurations for nodes with the Lagrangian flag. Results of the verification of these new procedure algorithms using two test problems with strong deformations of Lagrangian lines are presented. Keywords: artificial neural network, multilayer perceptron, maintenance of the spatial grid quality, gas dynamics.
|
English
