Session: 34-01 AI for flow field prediction and post-processing

Submission Number: 176797

Data-Driven Flow Field Prediction: Rules for Transparent Reporting 

Machine learning methods for flow field prediction are becoming more common, but reporting is often inconsistent. Differences in data usage, training conditions, model evaluation, and error analysis make it difficult to compare results, understand model limitations, or judge progress in the field. To address this, we propose a set of rules for developing, evaluating, and reporting models. The rules cover data usage, generality of training and test conditions, error analysis, interpolation and extrapolation behaviour, computational cost, and transparency about model limitations. Applying these rules can improve reproducibility, allow fair comparisons, and provide clearer guidance for future work.

We demonstrate the rules through an example that reconstructs two-dimensional flow fields for turbine and compressor blade sections. The method uses inverse design parameters as inputs, predicting both blade geometry and flow field simultaneously. The flow field prediction applies a two-step prediction: a spatial mapping followed by a scalar field prediction on the mapped space. Using this example, we show how each rule can be applied in practice, from defining training coverage and reporting field-level errors to evaluating performance on unseen conditions. The demonstration highlights how consistent reporting supports understanding of model behaviour, identifies failure modes, and sets a standard for more transparent and reliable development in data-driven flow field prediction.

 

Presenting Author: Christopher Clark Whittle Laboratory

Presenting Author Biography: Assistant Professor in propulsion and power at the Whittle Laboratory.

Authors:

Christopher Clark Whittle Laboratory