⚡ electrical engineering

Multi-Fidelity Surrogate Modelling for Uncertainty-Aware Prediction of Disturbed Velocity Fields in Virtual Flow Metering

This paper presents an auto-regressive multi-fidelity Gaussian process regression framework that fuses sparse, high-fidelity LDV measurements with abundant, biased CFD simulations to create a real-time, uncertainty-aware surrogate model for predicting disturbed velocity fields in virtual flow metering, achieving significant error reduction and reliable 95% prediction intervals.

Original authors: Nursen Bayazit, Martin Straka, Kilian Oberleithner, Sonja Schmelter

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nursen Bayazit, Martin Straka, Kilian Oberleithner, Sonja Schmelter

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the hidden world of industrial piping, where oil, gas, and water move through vast networks, the accuracy of measurement is a matter of safety, money, and environmental stewardship. To know exactly how much fluid is flowing, engineers rely on flow meters. However, these devices are typically calibrated in perfect, straight pipes where the fluid moves in a smooth, predictable stream. In the real world, pipes twist, turn, and connect in complex ways. These bends and junctions disturb the flow, creating swirls and uneven speeds that can trick a meter into giving the wrong reading. For decades, fixing this problem meant physically installing meters in every possible pipe configuration to test them, a process that is slow, expensive, and often impossible to do for every unique setup. Alternatively, scientists use powerful computer simulations to model these flows, but these digital models are not perfect; they contain small, systematic errors that can accumulate and mislead. The challenge has been to find a way to get the speed and coverage of computer simulations while keeping the absolute accuracy of real-world measurements, all while knowing exactly how much trust to place in the result.

A team of researchers at the Physikalisch-Technische Bundesanstalt and Technische Universität Berlin has developed a new method to solve this puzzle. They created a digital tool that acts as a bridge between the abundance of computer data and the scarcity of real measurements. The researchers focused on a specific, tricky scenario: a pipe with two S-shaped bends in a row. This configuration creates a complex, disturbed flow field that is difficult to predict. To build their solution, they gathered two types of information. First, they ran hundreds of detailed computer simulations of the fluid moving through these pipes. These simulations provided a rich, complete picture of the flow but carried a known bias, meaning they were consistently slightly off in specific ways. Second, they took a much smaller set of highly precise measurements from a real pipe using a laser-based technique that tracks the speed of the fluid at specific points. These real-world data points were sparse, covering only a fraction of the possible conditions, but they were metrologically traceable, meaning their accuracy was guaranteed by rigorous standards.

The core of their work was a sophisticated mathematical framework that learned how to correct the computer simulations using the real measurements. Instead of trying to predict the entire flow field at once, which would be computationally overwhelming, the researchers broke the flow down into simpler parts. They separated the main, symmetrical flow from the asymmetrical swirls and disturbances caused by the bends. They then used a statistical learning method to train a model that could take the abundant, slightly flawed computer data and adjust it using the sparse, highly accurate laser data. This process allowed them to generate a complete, two-dimensional map of the fluid's speed across the entire pipe cross-section, not just at the points where measurements were taken. Crucially, the model did not just give a single number; it calculated a range of uncertainty for every prediction, telling engineers exactly how much confidence they could have in the result.

The results of this approach were striking. When the researchers tested their new model against independent real-world measurements that were not used during the training process, the model reduced the error in predicting the fluid speed by more than 70 percent compared to the raw computer simulations. In some cases, the improvement was even higher, reaching over 80 percent. The model was able to predict the complex flow patterns with a high degree of precision, achieving an average error of less than 1.6 percent. Perhaps most importantly, the uncertainty estimates provided by the model were reliable. When the researchers checked how often the real measurements fell within the predicted range of uncertainty, they found that the model was correct 96.8 percent of the time, slightly better than the 95 percent target they had set. This means the tool does not just guess; it knows when it is guessing and provides a safety margin that engineers can trust.

The researchers also demonstrated that this method works in real-time. While running a single high-fidelity computer simulation of such a flow might take hours, their trained model could generate a prediction in less than half a second. This speed opens the door for rapid testing of countless pipe configurations without the need for expensive physical experiments. The model successfully handled different pipe curvatures, flow speeds, and distances between the bends, generalizing from the limited data it was given to cover a wide design space. However, the authors are careful to note that the model's performance is best where it has been trained with real data. In areas far from the measured conditions, the uncertainty naturally increases, reflecting the limits of the available information. The tool is not a magic replacement for all physical testing, but rather a powerful, uncertainty-aware assistant that makes virtual testing far more reliable.

This work represents a significant step forward in the field of virtual flow metering. By fusing the strengths of computer simulations and physical measurements, the researchers have created a system that can predict flow disturbances with a level of accuracy and reliability previously unattainable. The ability to quantify the uncertainty of these predictions is particularly vital for industries where measurement errors can have serious consequences. The method provides a foundation for establishing a new kind of metrological traceability, where the accuracy of a flow meter in a complex, disturbed pipe can be verified through a virtual experiment that is as trustworthy as a physical one. As the researchers look to the future, they aim to integrate this model directly into the calibration procedures for flow meters, ensuring that the digital twins of our industrial infrastructure are not just fast, but also rigorously accurate.

Technical Summary: Multi-Fidelity Surrogate Modelling for Uncertainty-Aware Prediction of Disturbed Velocity Fields in Virtual Flow Metering

Problem Statement
Accurate prediction of velocity fields under disturbed flow conditions is critical for quantifying measurement errors in flow meters and establishing metrological traceability in industrial piping systems. While Computational Fluid Dynamics (CFD) offers extensive coverage of geometrical configurations and operating conditions, it introduces systematic biases due to modelling and numerical simplifications. Conversely, Laser Doppler Velocimetry (LDV) provides highly accurate, traceable point measurements but is limited in spatial coverage and expensive to deploy for full-field characterization. Existing surrogate models for virtual flow metering primarily focus on scalar outputs (e.g., flow rates) and fail to provide spatially resolved predictions of the two-dimensional velocity field with calibrated uncertainty bounds. Furthermore, the systematic integration of abundant but biased CFD data with sparse but accurate LDV measurements into a unified framework for velocity field prediction remains an unaddressed gap.

Methodology
The authors propose a three-stage framework based on Auto-Regressive Multi-Fidelity Gaussian Process Regression (AR-MF GPR) to fuse CFD and LDV data:

Spectral Decomposition and Dimensionality Reduction:
The high-dimensional output space (1200 grid points per cross-section) is reduced by decomposing the normalized axial velocity field via azimuthal Fourier analysis. The field is separated into an axisymmetric base profile (mode $m=0$ ) and higher-order asymmetric disturbance modes ( $m \geq 1$ ). Using an energy threshold criterion, the study retains $K=5$ modes, capturing 95% of the disturbance energy. This reduces the feature dimension from 1200 to 330 spectral coefficients per sample, enabling computational efficiency while preserving the dominant flow structures, including secondary vortex pairs.
AR-MF GPR Modelling:
Separate multi-fidelity Gaussian Processes are trained for the axisymmetric and asymmetric components. The framework utilizes 196 CFD simulations as the low-fidelity source and 28 LDV measurements as the high-fidelity reference.
- Low-Fidelity GP ( $G_{P,L}$ ): Trained on all CFD data to capture global trends and geometric dependencies.
- Discrepancy GP ( $G_{P,D}$ ): Trained on the residuals between the scaled CFD predictions and LDV measurements to correct systematic biases.
- Scaling Factor: A global scaling factor ( $\rho$ ) is estimated via least squares to align the amplitude of the low-fidelity predictions with high-fidelity data before discrepancy correction.
- Input Space: The model maps a four-dimensional feature vector (curvature radius $R_c/D$ , volume-averaged velocity $w_{vol}$ , downstream distance $z/D$ , and inter-bend distance $d_l/D$ ) to the spectral coefficients.
Conformal Calibration of Uncertainty:
To ensure reliable uncertainty quantification independent of the GP's distributional assumptions, the posterior variances are conformally calibrated. Using a held-out calibration set, a calibration factor ( $k_{conf}$ ) is determined to guarantee that the prediction intervals achieve a specific empirical coverage probability (95%). This step corrects potential over- or under-confidence in the raw GP posterior variances.

Key Contributions

Unified Multi-Fidelity Framework: The paper presents a novel approach to fuse sparse, metrologically traceable LDV data with abundant, physically consistent but biased CFD data for full-field velocity prediction.
Spatially Resolved Uncertainty: Unlike previous surrogate models focusing on scalar quantities, this framework predicts the entire two-dimensional velocity distribution across pipe cross-sections with calibrated uncertainty bounds.
Dimensionality Reduction via Spectral Analysis: The use of azimuthal Fourier decomposition allows for efficient training of multi-output GPs by reducing the output space while maintaining physical fidelity of the flow structures.
Distribution-Free Calibration: The application of conformal prediction ensures that the reported uncertainty intervals have a guaranteed empirical coverage, addressing the miscalibration often found in standard GP posterior intervals.

Results
The model was validated against independent LDV data not used in training or calibration:

Accuracy: The AR-MF GPR reduced the Root Mean Square Error (RMSE) of the normalized axial velocity by 72.9% at the near-disturbance position ( $z/D = 2.43$ ) and 82.1% at the far-field position ( $z/D = 15.69$ ) compared to uncorrected CFD.
Bias: The model achieved a near-zero mean bias ($-0.00005$) on the validation set, indicating no systematic over- or under-prediction.
Uncertainty Quantification: The conformally calibrated 95% prediction intervals achieved an empirical coverage of 96.8%, demonstrating that the uncertainty estimates are well-calibrated.
Performance: The trained model predicts in real-time (less than 0.5 seconds per evaluation), making it suitable for rapid parametric studies, whereas the underlying CFD simulations require hours of computation.

Significance and Claims
The paper claims that this framework enables uncertainty-aware virtual flow metering. By providing a surrogate model that predicts the full velocity field with calibrated uncertainties, it allows for the direct computation of installation-effect corrections for any flow meter type without requiring new physical experiments for every configuration. The authors emphasize that the model is particularly valuable in regions of the design space where no experimental data exists, as it leverages the geometric trends learned from CFD while correcting systematic biases using sparse high-fidelity data. The work establishes a foundation for propagating velocity field uncertainties through the complete virtual measurement chain to determine the uncertainties of fluid mechanical calibration factors, thereby supporting metrological traceability in industrial piping systems.

The authors modestly note limitations, including the assumption of spatially homogeneous noise, the neglect of correlations between Fourier coefficients, and the current restriction to a specific double S-elbow configuration. Future work is identified as extending the CFD database to additional Reynolds numbers and integrating the surrogate directly into virtual flow meter calibration procedures.

Technical Summary: Multi-Fidelity Surrogate Modelling for Uncertainty-Aware Prediction of Disturbed Velocity Fields in Virtual Flow Metering

More like this