Effective cooling is critical for the optimal performance and longevity of electrical machines. This paper presents an approach to improve the estimation of heat transfer coefficients (HTC) in air-fin cooled electrical machines featuring stator flux barriers. A Generalized k-omega (GEKO) turbulence model is tuned using machine learning (ML) and adjoint optimization, to allow high-fidelity prediction in complex geometries at the cost of a Reynolds-Averaged Navier-Stokes (RANS) simulation. An initial field of boundary conditions, estimated via conjugate heat transfer (CHT) simulation, is mapped to a Large-Eddy Simulation (LES) as the benchmark. Subsequently, a RANS-GEKO simulation, coupled with adjoint optimization and ML, is conducted to tune the turbulence coefficients. Firstly, to validate the procedure against LES and test data, the tuned GEKO model is applied to a geometry from the literature: a rectangular section channel with transverse periodic ribs. Relative to the original RANS model, the tuned GEKO model demonstrates an increased local HTC in the vicinity of the fins, decreasing HTC deviations from 25% to just 5.0% relative to the LES and to test data. This tuning is primarily attributed to the element-wise enhancement of the turbulent kinetic energy (TKE) parameter C k during ML, resulting in a local increase in TKE of up to 59%. Secondly, to guide the design of a novel stator flux barrier cooling solution for an annular electrical machine, the GEKO model is re-tuned. In this application, the ML correction of tuning parameters for TKE and specific dissipation rate (SDR) is impactful. The relative HTC difference is reduced from 32% to 7.5% relative to the LES results, outperforming standard RANS and substantiating its capability of high-fidelity CHT calculation. Lastly, the re-tuned model is used for an optimization study with respect to their rib separation, rib thickness, and flux barrier gap. Results of maximal stator temperatures range between 115 °C and 132 °C, while indicating a preference towards thinner ribs with relatively small separation spacing and a narrow flux barrier gap. • Tuning a Generalized k-ω (GEKO) turbulence model for conjugate heat transfer via adjoint optimization and machine learning. • Up to 59% enhancement of turbulent kinetic energy in a ribbed channel domain, reducing heat transfer error from 25% to 5%. • Tuned model is validated against LES and test data, providing high-fidelity results at the cost of a RANS simulation. • Validated model is used for a design study, to optimize stator flux barriers in an electrical machine. • Maximal stator temperature is reduced with thin ribs at small separation spacing and a narrow flux barrier gap.
Stiehl et al. (Thu,) studied this question.