Machine learning aided simulation of gas distribution systems operating under any vacuum conditions

Abstract

Gas distribution systems operating under any vacuum conditions are critical in technologies such as semiconductor manufacturing, microfluidics, vacuum metallurgy and nuclear fusion. Typical hydrodynamic models fail to predict gas behavior accurately across the wide range of rarefaction conditions encountered in these systems, necessitating computationally intensive mesoscale kinetic modeling. This type of networks may be efficiently simulated by in-house codes, such as ARIADNE, which combines a network solver with a kinetic database of the flow rates of rarefied gas flow through capillaries.

Here, ARIADNE is further enhanced by substituting the flow rates kinetic database with corresponding closed-form expressions, obtained by machine learning techniques, namely symbolic regression (SR). SR expressions, which have been deduced for flows through single pipe elements, are directly incorporated into the network solver and the need of the kinetic database is eliminated. The upgraded code is successfully validated by solving three benchmark gas distribution systems: two moderate scale networks and a very large one representing the primary vacuum system of the ITER fusion reactor. The presented results demonstrate the capability and robustness of the enhanced ARIADΝE code, named ARIADNE-ML, to accurately simulate vacuum systems of arbitrary size and complexity. Since there is no data dependency, the solver becomes more flexible, applicable and scalable across more complex systems without worrying about query bottlenecks, broadening its use in engineering and industrial applications.