Machine learning benchmark for flow reconstruction in the TCC–III optical engine

We present EngineBench, the first machine learning (ML) benchmark designed for engine in-cylinder flow research. The benchmark data consist of curated particle image velocimetry (PIV) measurements previously gathered from the Transparent Combustion Chamber (TCC-III) by the General Motors University of Michigan Automotive Cooperative Research Laboratory. 1 The dataset is then leveraged in order to benchmark the performance of four ML methods for a flow reconstruction (inpainting) task. We propose large gaps at the edges of the field of view as the benchmark task in order to reflect realistic scenarios in which data are harder to obtain closer to walls, and to challenge the ability of the models to predict the turbulent flow motion with limited access to surrounding data points. Pixel-wise, vector-based and multi-scale performance metrics are used to provide broad evaluations of the models. We find that models which utilise skip connections show significantly improved performances at this task on both small and large gap sizes, due to their enhanced ability to leverage contextual information. The benchmark proposed in this paper supports the development of ML models for engine design problems, as well as PIV data enhancement more generally. In addition, the ML model comparisons allow for more informed selection of models for problems in experimental flow diagnostics. All data and code are publicly available at https://eng.ox.ac.uk/tpsrg/research/enginebench /.