Flow based features and validation metric for machine learning reconstruction of PIV data

TL;DR

This paper introduces flow-based features and a mass conservation validation metric to improve machine learning reconstruction of sparse PIV flow data, demonstrating high accuracy with SVR and MLP models.

Contribution

It proposes a physics-oriented feature extraction and a mass conservation metric for better flow data reconstruction using machine learning.

Findings

SVR achieved R2-score of 0.993 for velocity reconstruction.

SVR outperformed MLP in mass conservation accuracy.

Reconstructed flow fields with 75% missing data remained consistent with original data.

Abstract

Reconstruction of flow field from real sparse data by a physics-oriented approach is a current challenge for fluid scientists in the AI community. The problem includes feature recognition and implementation of AI algorithms that link data to a physical feature space in order to produce reconstructed data. The present article applies machine learning approach to study contribution of different flow-based features with practical fluid mechanics applications for reconstruction of the missing data of turbomachinery PIV measurements. Support vector regression (SVR) and multi-layer perceptron (MLP) are selected as two robust regressors capable of modelling non-linear fluid flow phenomena. The proposed flow-based features are optimally scaled and filtered to extract the best configuration. In addition to conventional data-based validation of the regressors, a metric is proposed that reflects mass conservation law as an important requirement for a physical flow reproduction. For a velocity field including 25% of clustered missing data, the reconstruction accuracy achieved by SVR in terms of R2-score is as high as 0.993 for the in-plane velocity vectors in comparison with that obtained by MLP which is up to 0.981. In terms of mass conservation metric, the SVR model by R2-score up to 0.96 is considerably more accurate than the MLP estimator. For extremely sparse data with a gappiness of 75%, vector and contour plots from SVR and MLP were consistent with those of the original field.