Data-Driven Flow Field Prediction: Rules for Transparent Reporting

Machine learning (ML) methods are increasingly applied to predict integral properties and flow fields in turbomachinery, offering the potential to drastically reduce computa tional cost compared with traditional CFD solvers. Despite this growing interest, there is little consensus on best practices for reporting and evaluating surrogate models. Many studies focus solely on cases where the surrogate performs well, providing an incomplete picture of model reliability and limitations. This paper provides a set of transparent rules of engagement for the evaluation and reporting of ML-based flow field prediction. The guidelines cover data usage, model speed, interpolation and extrapolation behavior, error analysis, and failure modes, calling for clearer and more complete reporting to accelerate progress in the sector. To illustrate the application of these guidelines, a worked example is presented using a polynomial regression surrogate trained on 2400 inverse-design-generated blade sections. Predictions include geometry, 2D flow fields, boundary-layer quantities, and performance metrics. Accuracy is assessed across interpolation and high-dimensional extrapolation tests, showing that errors increase with distance from the training manifold, but remain moderate due to the simple underlying physics captured by inverse parameterization. The paper demonstrates how to report both successes and limitations, including failure cases for shocks, extreme angles, and diffusion. These results provide practical recommen dations for surrogate model evaluation, ensuring transparency, reproducibility, and safe deployment in design workflows.