Central to data science is machine learning, which is a set of algorithms that allows systems to automatically learn directly from data by finding relations between inputs, outputs and parameters. Machine-learning algorithms have greatly advanced thanks to step changes in computer hardware, efficient algorithms, exa-scale amounts of data, and high-performance computing. Fluid mechanics is one of the original big-data communities. The fluid-mechanics community has been using data-driven and machine-learning techniques to guide large-scale simulations, interpret experimental data, and derive reduced-order models. Examples in fluids are: flow-feature extraction for reduced-order modelling; dimensionality reduction; classifications of wake topology; sparse compressed sensing for wall-bounded turbulence; trajectory analysis and classification of particle-image velocimetry; reconstruction of turbulent flow fields; identification of coherent structures from time-series data; super-resolution of flow fields; flow control; and many other applications, for example, in reinforcement learning and sparse identification. These-machine learning techniques have been applied to benchmark problems with success, but some questions are still open: (i) Do data-driven and machine-learning tools scale to engineering configurations? (ii) How can we gain physical insight and causal relations into the solutions? (iii) Can we extrapolate knowledge to other configurations? The objectives of this workshop are:

• bring fluid dynamicists together to address these questions;
• discuss the emergence of data-driven methods, machine learning and optimization in fluid mechanics;
• identify challenges to address and establish open datasets for training and benchmarking.

This workshop will bring together communities from turbulence and dynamical systems theory that use data-driven methods. The topics that will be included are:

• Physics-aware machine learning

The session will have a keynote speaker (potentially G Karniadakis) followed by talks with the traditional duration of EuroMech colloquia. The keynote will also be asked to prepare a summary of “do’s” and “don’ts” in physics-informed machine learning. We will particularly welcome submissions in subtopics: Solutions of inverse problems; solutions of Navier-Stokes; closure modelling (RANS and LES); flow reconstruction; etc.

• Data assimilation
  

The session will have a keynote speaker (potentially G Iaccarino) followed by talks with the traditional duration of EuroMech colloquia. The keynote will be asked to prepare a summary of “do’s” and “don’ts” in data assimilation. We will particularly welcome submissions in subtopics: Bayesian methods for real-time data assimilation; variational methods for offline assimilation; inference of model error; assimilation of experimental data; mean flow reconstruction; applications in thermoacoustics and non-reacting flows; etc.

• Gaining insight into flow physics through machine learning

The session will have a keynote speaker (potentially M Brenner) followed by talks with the traditional duration of EuroMech colloquia. We will particularly welcome submissions in subtopics: Nonmodal and nonlinear decomposition (generalization of POD, DMD); sensitivity of turbulent flows; flow structures before and during extreme events (such as relaminarization events); etc.

• Reduced-order modelling

The session will have a keynote speaker (potentially T Colonius) followed by talks with the traditional duration of EuroMech colloquia. We will particularly welcome submissions in subtopics: Learning the latent space and attractor of prototypical turbulent problems; modelling atomization with autoencoders; flow reconstruction from sparse sensing; etc.