A physics-informed neural network for supersonic strut-injector flowfield prediction under varying Mach numbers

Abstract

Rapid flowfield prediction is essential for scramjet-related hydrogen injection and mixing under varying Mach numbers. This study develops a physics-informed neural network for a Deutsches Zentrum für Luft- und Raumfahrt-type supersonic strut-injector configuration under non-reacting hydrogen–air mixing conditions. Validated two-dimensional Reynolds-averaged Navier–Stokes simulations are used to generate reference data. Taking spatial coordinates and incoming Mach number as inputs, the network predicts pressure, temperature, velocity components, and species mass fractions. Continuity and momentum residuals are combined with the supervised loss in the composite training objective to improve physical consistency. A Schlieren-informed stratified weighted sampling strategy further emphasizes shock-dominated and high-gradient regions. Results show that partial differential equation constraints enhance extrapolative robustness, suppress nonphysical artifacts, and better preserve flowfield morphology under unseen Mach numbers. The proposed method provides an effective physics-constrained surrogate for complex supersonic mixing flows.