PaperlayerView exact revision on arXiv

Source-linked AI summary

The transformative potential of machine learning for experiments in fluid mechanics

Ricardo Vinuesa, Steven L. Brunton, Beverley J. McKeon

arXiv:2303.15832v2physics.flu-dyncs.AI

TL;DR

Experiments are important for observing real-world flows, but non-ideal conditions and diagnostic limits constrain what they can measure. This perspective surveys ML approaches for measurement enhancement, experimental design and surrogate modeling, and real-time estimation and control. It describes immediate benefits for improving PIV velocity fields and active learning for efficient parameter-space sampling.

  • Problem

    Experimental observations of real-world flows are valuable, but noise, drift, and diagnostic limits such as restricted resolution and sensor bias complicate measurements.

    Source: Improvement of measurement techniques… · p. 2

  • Method

    The perspective surveys ML applications for measurement fidelity, experimental design and surrogate modeling, and real-time estimation and control.

    Source: Introduction · p. 1, Conclusions and outlook · p. 7

  • Results

    ML techniques have shown immediate benefit in improving PIV velocity fields, while active learning offers efficient sampling of parameter spaces with error bounds and uncertainty estimates for resulting models.

    Source: Conclusions and outlook · p. 8

  • Takeaways & Limitations

    Image-processing ML can improve measurement capabilities without more costly experimental equipment by leveraging data from high-fidelity measurements and computations.

    Source: Conclusions and outlook · p. 8

    Real-time flow control remains challenging with many sensors and actuators, high-dimensional search spaces, strong nonlinearities, and time delays.

    Source: Conclusions and outlook · p. 8

Abstract

from arXiv · show

The field of machine learning has rapidly advanced the state of the art in many fields of science and engineering, including experimental fluid dynamics, which is one of the original big-data disciplines. This perspective will highlight several aspects of experimental fluid mechanics that stand to benefit from progress advances in machine learning, including: 1) augmenting the fidelity and quality of measurement techniques, 2) improving experimental design and surrogate digital-twin models and 3) enabling real-time estimation and control. In each case, we discuss recent success stories and ongoing challenges, along with caveats and limitations, and outline the potential for new avenues of ML-augmented and ML-enabled experimental fluid mechanics.

Introduction

Experiments reveal real-world flow behavior and can be essential when flows reach conditions beyond numerical development, but non-ideal conditions and diagnostic limits complicate measurements. The paper organizes ML applications in experimental fluid mechanics into measurement enhancement, experimental design and modeling, and estimation and control.

Introduction

  • Experiments can be the first or only route to observing some flows and can complement, validate, or extend theory and simulation.Noise, vibration, thermal drift, and limits in spatial or temporal resolution, sensor accuracy, and bias complicate those observations.Source: Introduction · p. 1
  • The paper groups ML applications into measurement-fidelity enhancement, experimental design and modeling, and estimation and control.The categories are ordered by the increasing novelty of ML-enabled capabilities to the field.Source: Introduction · p. 1
  • Figure 1 classifies example tasks and ML improvements across those categories, with arrows indicating possible flows and rankings illustrating ML and experimental complexity.The classification is illustrative, not exhaustive or a categorical judgment of method merit; complexity also depends on how the ML method is implemented in an experimental task.Source: Introduction · pp. 1–2, Figure 1 · p. 2

Improvement of measurement techniques by means of machine learning

ML can augment measurement techniques by learning corrections for sensor imperfections and extracting information absent from measurements. The section describes approaches spanning sensor correction, visualization, and PIV enhancement.

Improvement of measurement techniques by means of machine learning

  • ML may learn sensor transfer functions to address bias or noise, while learned corrections could replace empirical forms that hinder general use.Pitot-tube corrections include viscous effects, shear, wall interference, and turbulence; for hot-wire filtering, small-scale non-universality under pressure gradients limits correction schemes.Source: Improvement of measurement techniques… · p. 2
  • CNNs trained on matched simulations could map smoke-wire visualizations to velocity and pressure vectors, adding quantitative information to qualitative flow images.Deep-learning approaches can first enhance visualization resolution and then add quantitative information.Source: Improvement of measurement techniques… · p. 3
  • PIV enhancement approaches include POD noise removal, RPCA outlier removal and occlusion filling, and neural-network post-processing and super-resolution.CNN-based autoencoders have also been used for robust flow estimation.Source: Improvement of measurement techniques… · p. 3
  • PINNs can improve noisy measured flow fields [29] and PIV resolution and quality in turbulence [119], while vortex-in-cell methods can improve PIV temporal resolution.PINNs train neural networks by minimizing PDE residuals, enabling theoretical considerations to be embedded in the solution.Source: Improvement of measurement techniques… · p. 3

Experimental design and modeling enhanced by machine learning

ML-supported experimental design, reduced-order modeling, uncertainty quantification, and digital twins can help characterize and optimize fluid experiments. These approaches combine measurements with data-driven and physics-based models, while active learning can guide sampling.

Experimental design and modeling enhanced by machine learning

  • Gaussian-process optimization of wind-tunnel inserts can converge in few iterations while selecting the next parameter case to explore, but is suited only to relatively few parameters.Insert shape is optimized to achieve a desired pressure-gradient distribution when wind-tunnel conditions are difficult to realize.Source: Experimental design and modeling… · pp. 3–4
  • Reduced-order models can predict, estimate, and control flows and apparatus; sparse nonlinear models have shown promise for interpretable, generalizable models from experimental data.A parsimonious model for an actuated D-shaped body captured resonance and lock-on phenomena.Source: Experimental design and modeling… · p. 4
  • Comprehensive UQ methods accounting for correlations among hot-wire measurements yielded lower uncertainties than propagation of errors because of error cancellation.For hot-wire measurements, wire voltage had the highest impact; for oil-film interferometry wall-shear stress, fringe-velocity calculation had the largest effect.Source: Experimental design and modeling… · p. 5
  • A digital twin combines physical, machine-learned, and hybrid models, updated by real-world data, and may support optimization, iterative design, and control.Active learning and experimental design help sample the system, while a hierarchy of physics-based and data-driven models can characterize the twin.Source: Experimental design and modeling… · pp. 3, 5, Figure 3 · p. 5

Possibilities within data-driven flow estimation and control

Data-driven methods use sparse wall measurements to estimate flow and support sensing and control, while fluid flows’ high dimensionality, nonlinearity, fast timescales, instabilities, and delays complicate implementation.

Conclusions and outlook

The review closes by identifying ML opportunities in measurement fidelity, experimental design and digital-twin surrogates, and real-time estimation and control. It also notes open questions in digital-twin pipelines and challenges for flow control.

Conclusions and outlook

  • ML can improve PIV velocity fields through super-resolution, outlier detection, and imputation; physics constraints improve consistency from limited measurements without costly new equipment.These efforts can leverage data from other high-fidelity measurements and computations.Source: Conclusions and outlook · pp. 7–8
  • Active learning can sample experimental parameter spaces efficiently with error bounds and uncertainty estimates, supporting surrogate models for iterative design and downstream estimation and control.Digital twins combine physics-based, machine-learned, or hybrid model hierarchies pinned to measurement data with uncertainty quantification; their pipelines remain under development.Source: Conclusions and outlook · p. 8
  • High-dimensional sensor-actuator spaces, nonlinearities, and time delays challenge real-time flow control, while reinforcement learning and ML-enabled model predictive control are promising.When appropriate, simpler strategies such as simplex optimization may also be advantageous; it improved power efficiency by over 50% in a lab-scale cross-flow turbine.Source: Conclusions and outlook · p. 8
  • Table 1 offers experimentalists a selection of sample ML methods to help choose an approach for an intended experimental-fluid-mechanics task.The table marks experimental techniques or tasks in italics and relevant ML methods in bold.Source: Conclusions and outlook · p. 8, Table 1 · p. 8
  • Cloud connectivity may enable remotely controlled, collaborative fluid-flow facilities and a new generation of shared experimental resources.Source: Conclusions and outlook · p. 9
  • Experiments remain the gold standard for capturing multiscale, multiphysics effects in complex-geometry flows, while transferring results across configurations remains an open problem.The authors expect these approaches to mature rapidly, with aerospace and biomedical applications identified as early adopters.Source: Conclusions and outlook · p. 9
Loading 2303.15832v2…