Enhancing vortex-flow-meter precision using physics-informed contrastive learning

High precision vortex-based flow measurement devices are subject to systematic measurement errors caused by installation effects and complex flow conditions that cannot always be directly measured or compensated. Correcting such systematic measurement errors is crucial for achieving a yet higher measurement accuracy and reliability. In this work we introduce a hybrid framework for error correction of vortex flow meters. The method uses physically-engineered features derived from computational fluid dynamics (CFD) simulations as inputs to a deep contrastive regression neural network. The network learns latent representations that are useful for predicting measurement errors under various pipe geometries and flow regimes. We demonstrate the effectiveness of this representation learning by testing the prediction performance of the framework under new pipe geometries not seen during training. The method demonstrates the potential of advanced deep learning models to extract physically meaningful features for error prediction tasks in complex, highly non-linear flow setups, in which CFD simulations reach their computational limits. • Hybrid CFD and deep learning for flow sensor error correction • A first practically viable calibration framework for vortex-based flow meters • Out-of-distribution generalization for new installations and operational conditions • Contrastive regression for technical images