Source-linked AI summary
Applying Machine Learning to Study Fluid Mechanics
Steven L. Brunton
TL;DR
Fluid mechanics has many data-rich problems suited to machine learning, but creating useful models requires expert choices rather than a turn-key process. This paper surveys five stages of model building and how physical knowledge can be incorporated at each. It concludes that physics-informed modeling is a collaborative effort, with physics framed as what generalizes in a model.
Problem
Fluid mechanics presents nonlinear, high-dimensional modeling and optimization problems, while building models from data requires expert guidance rather than a turn-key procedure.
Source: §1 · p. 1Method
The paper gives a high-level overview of five machine-learning stages and ways to incorporate physical knowledge, illustrated with fluid-mechanics examples.
Source: §2 · p. 3, §2.1 · p. 4Results
The paper concludes that each stage offers an opportunity to incorporate physical knowledge, and that machine learning for physical systems is fundamentally collaborative.
Source: §3 · p. 8Takeaways & Limitations
The authors frame physics as the part of a model that generalizes and encourage revisiting classical problems where complex-function representation limited progress.
Source: §3 · p. 8The examples are not exhaustive, provide only a small glimpse of the field, and disproportionately cite the authors’ close collaborators’ work.
Source: Disclaimer · p. 9
Abstract
from arXiv · showhide
This paper provides a short overview of how to use machine learning to build data-driven models in fluid mechanics. The process of machine learning is broken down into five stages: (1) formulating a problem to model, (2) collecting and curating training data to inform the model, (3) choosing an architecture with which to represent the model, (4) designing a loss function to assess the performance of the model, and (5) selecting and implementing an optimization algorithm to train the model. At each stage, we discuss how prior physical knowledge may be embedding into the process, with specific examples from the field of fluid mechanics.
1 Introduction
Machine learning fits fluid mechanics’ nonlinear, high-dimensional optimization problems, but building useful models requires expert guidance throughout the process. The overview frames physics-informed choices as a way to improve generalization and interpretability amid both optimism and skepticism.
1 Introduction
- Fluid mechanics offers nonlinear, high-dimensional optimization problems, from estimating flows with limited measurements to designing wings or controlling turbulence.Machine learning can either provide a surrogate to optimize or directly address an engineering objective through active control.Source: §1 · p. 1
- Machine learning is not a turn-key route from data to models; experts must guide problem choice, data curation, architecture, loss design, and optimization.Models generally fit into established model-based design, optimization, and control rather than acting as general artificial intelligence.Source: §1 · pp. 1–2
- Physics-informed choices in the learning pipeline are presented as a route to better generalization, interpretability, and explainability.The paper notes that training a model for a defined task is easier than creating one that outperforms traditional numerical algorithms and physics-based models.Source: §1 · p. 2
2 Physics Informed Machine Learning for Fluid Mechanics
The paper organizes applied machine learning into five connected stages, each offering opportunities to incorporate physical knowledge. It surveys how those choices shape models and applications in fluid mechanics.
2 Physics Informed Machine Learning for Fluid Mechanics
- The five-stage workflow covers problem choice, data, architecture, loss, and optimization, with physics knowledge incorporable throughout and later outcomes prompting revisions.The stages overlap: problem and data choices are related, while loss design and optimization are tightly coupled.Source: §2 · pp. 2–3
2.1 The problem
- A modeling problem specifies available inputs and desired outputs; that choice distinguishes supervised, unsupervised, and reinforcement-learning tasks.Fluid-mechanics examples include turbulence closure, solver improvement, super-resolution, modal decomposition, control, and experiment design, often within physics-based frameworks.Source: §2.1 · pp. 3–4
2.2 The data
- Training-data quality and quantity affect model performance, while physical knowledge can inform augmentation, feature design, and use of multifidelity sources.Flow data can be vast in space and time yet sparse across parameters such as geometry and Reynolds number, because simulations and experiments are costly.Source: §2.2 · pp. 4–5
2.3 The architecture
- An architecture defines the function family used to represent a model, and physical structure can be built into architectures to guide what they learn.Examples include symmetry-aware networks, autoencoders that impose a low-dimensional bottleneck, physics-informed networks constrained by governing laws, and operator networks.Source: §2.3 · pp. 5–6
2.4 The loss function
- Loss functions quantify performance and can balance fit against complexity while encoding physical priors such as sparsity, stability, or energy conservation.Fluid applications include sparse reduced-order and closure models, bounded-solution losses, and divergence penalties that promote incompressibility.Source: §2.4 · pp. 6–7
2.5 The optimization algorithm
- Training is difficult because high-dimensional, non-convex optimization offers few general guarantees, so algorithms may need to accommodate physical constraints and custom losses.For incompressible flows, skew-symmetry constraints can enforce energy conservation in sparse regression; KKT conditions impose them during regression, and SR3 handles more challenging constrained problems.Source: §2.5 · pp. 7–8
3 Parting Thoughts
The paper offers a high-level, deliberately simplified overview of machine learning stages and their use in fluid mechanics. It frames physics as what generalizes in a model and encourages revisiting important problems with data-driven tools.
3 Parting Thoughts
- The overview treats prior physical knowledge as relevant at every stage of building machine-learning models for physical systems.It describes machine learning as collaborative work because mastering every stage is nearly impossible for one person.Source: §3 · p. 8
- The paper’s working definition of physics is the part of a model that generalizes, a central goal for machine-learning models of physical systems.Source: §3 · p. 8
- The authors recommend revisiting classical problems where progress was limited by the ability to represent complex functions.They cite Ling et al.’s work revisiting classical Reynolds stress models with modern techniques.Source: §3 · p. 8
Disclaimer
The author cautions that these notes are selective rather than exhaustive, and that their references disproportionately reflect work by close collaborators.
Disclaimer
- The examples offer a small, non-exhaustive glimpse of a rapidly growing field, not a comprehensive account.The author chose work considered relevant and accessible.Source: Disclaimer · p. 9
- References disproportionately cover the author's close collaborators because that work is most familiar to the author.Source: Disclaimer · p. 9
- The author invites corrections to missed references or mischaracterized work for possible incorporation in future versions.Source: Disclaimer · p. 9