EngineBench: Flow Reconstruction in the Transparent Combustion Chamber III Optical Engine

TL;DR

EngineBench provides the first real-PIV, ML-focused dataset for turbulent internal flows in propulsion, introducing a realistic random edge-gap inpainting benchmark. By comparing GPOD, UNet, UNETR, and CE-GAN on 10% and 25% edge gaps, the study shows UNet-based models deliver the best pixel- and structure-level reconstruction, with UNETR offering limited gains given dataset size. The work demonstrates that central flow features are well-captured, while edge gaps pose significant challenges, and it offers open data and code to drive reproducible development of industry-relevant ML models for flow diagnostics and digital twin applications.

Abstract

We present EngineBench, the first machine learning (ML) oriented database to use high quality experimental data for the study of turbulent flows inside combustion machinery. Prior datasets for ML in fluid mechanics are synthetic or use overly simplistic geometries. EngineBench is comprised of real-world particle image velocimetry (PIV) data that captures the turbulent airflow patterns in a specially-designed optical engine. However, in PIV data from internal flows, such as from engines, it is often challenging to achieve a full field of view and large occlusions can be present. In order to design optimal combustion systems, insight into the turbulent flows in these obscured areas is needed, which can be provided via inpainting models. Here we propose a novel inpainting task using random edge gaps, a technique that emphasises realism by introducing occlusions at random sizes and orientations at the edges of the PIV images. We test five ML methods on random edge gaps using pixel-wise, vector-based, and multi-scale performance metrics. We find that UNet-based models are more accurate than the industry-norm non-parametric approach and the context encoder at this task on both small and large gap sizes. The dataset and inpainting task presented in this paper support the development of more general-purpose pre-trained ML models for engine design problems. The method comparisons allow for more informed selection of ML models for problems in experimental flow diagnostics. All data and code are publicly available at https://eng.ox.ac.uk/tpsrg/research/enginebench/.

Figures (10)

• Figure 1: Example PIV image, showing a circular field of view. At each pixel the arrows show the direction of the turbulent flow, and the colourmap shows the velocity magnitude.
• Figure 2: Schematic showing the TCC-III and associated PIV measurement planes.
• Figure 3: Example random edge gap creation. From left to right, (a): original image; (b): image with a random edge gap polygon superimposed in red; (c): edge gaps added to regions outside of the random polygon.
• Figure 4: Histogram showing the proportion of pixels removed by the random edge masks in one pass through the training set. 7% of the total pixels in the field of view were removed on average.
• Figure 5: Energy spectra comparing the ground truth test set images to the UNet, MSE predictions at a 10% test gap size. Ensemble mean spectra are given by solid or dashed lines, with the shaded areas representing one standard deviation from the means.