IRMA: Impeller Representation Model for Aerodynamics

Accurate aerodynamic modeling requires representations capable of capturing both the smooth spatial variations of blade geometry and the complex behavior of flowfields. Impeller geometry varies continuously along chordwise and spanwise directions and contains detailed features that significantly affect performance. Flowfields, in contrast, can exhibit sharp gradients and localized discontinuities, such as shock waves in transonic conditions, which must be resolved to accurately represent aerodynamic behavior. Effectively linking geometric features with resulting aerodynamic response is therefore essential for reliable surrogate modeling and design exploration. This work introduces IRMA (Impeller Representation Model for Aerodynamics), a multimodal neural network that learns a shared latent representation capturing the mapping between three-dimensional impeller geometry and its aerodynamic response. The model simultaneously encodes discrete geometric parameters and image-based representations of blade-angle distributions, from which coupled decoders predict both CFD-derived flowfields and global performance metrics. By explicitly modeling the relationship between geometry and aerodynamic outputs, IRMA provides a compact and expressive surrogate capable of linking continuous design features to both spatially distributed and integral aerodynamic quantities, enabling efficient analysis and optimization of centrifugal impeller geometries.