Flow-Regime Classification Neural Network for IAUGKS in Supersonic Rarefied Flow

This study proposes a branched shape-aware flow regime classification neural network (SA-FCNet) and integrates it with the implicit adaptive unified gas-kinetic scheme (IAUGKS) to accelerate multiscale supersonic flow simulations. SA-FCNet introduces a three-branch structure that incorporates local flow features, global flow context, and local geometry, enabling improved physical awareness and geometric generalization. This design allows a network trained on one geometry to be applied to others without bias toward specific training cases. A physics-based classification label is defined using the relative deviation between the actual distribution (from a 1000-step implicit gas-kinetic scheme (IGKS) and a one-step implicit unified gas-kinetic scheme (IUGKS)) and the local Maxwellian, avoiding full convergence and reducing offline data cost. SA-FCNet achieves accurate regime prediction across a range of Mach and Knudsen numbers, from a circular cylinder to realistic two-dimensional vehicles with internal–external flows, and generalizes well to unseen geometries. The network achieves classification accuracy above 97% and geometric generalization accuracy above 91%, reaching over 96% in most cases. Compared to gradient-length local-Knudsen-number-based classifiers, it reduces misclassification in shock and wake regions. Coupled with IAUGKS, the method enables selective reduction of the discrete velocity space, preserves accuracy close to full IUGKS, and lowers [Formula: see text] computational cost.