A Physics-Informed Machine-Learning Platform for Smart Lagrangian Harness and Control of TURBulence

Projektinformationen

Smart-TURB

ID Finanzhilfevereinbarung: 882340

DOI

10.3030/882340

Projekt abgeschlossen

EK-Unterschriftsdatum 14 Mai 2020

Startdatum 1 Mai 2021

Enddatum 30 April 2026

Finanziert unter

• EXCELLENT SCIENCE - European Research Council (ERC)

Gesamtkosten

€ 2 248 875,00

EU-Beitrag

€ 2 248 875,00

Koordiniert durch

UNIVERSITA DEGLI STUDI DI ROMA TOR VERGATA
Italy

Periodic Reporting for period 3 - Smart-TURB (A Physics-Informed Machine-Learning Platform for Smart Lagrangian Harness and Control of TURBulence)

Berichtszeitraum: 2024-05-01 bis 2025-10-31

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

The project is attacking important issues connected to Lagrangian and Eulerian data-driven problems for complex flows, with particular interests to applications to geophysical and laboratory set-ups. We have started to attach single-probe self-propelled objects with specific point-to-point navigation targets in complex flows, and multi-probe cooperative or adversarial games in stochastic and/or turbulent flow Field-taxis based on optimal and heuristic search for odor emissions in stochastic models and in realistic numerical simulations of turbulence with a mean wind have also been attacked. Among the various techniques developed we cite the ones based on Reinforcement learning, Adversarial Reinforcement Learning and optimal control.
Data assimilation problems for rotating turbulent flows and lagrangian turbulent trajectories have been also developed by using both equations-based and data-driven tools, including linear methods based on principal orthogonal decomposition and non-linear ones, based on convolutional neural networks, Generative Adversarial Networks and, more recently, Diffusion Models. Comparison against Nudging is also systematically performed. Explorative studies of complex multi-component and multi-phase flows, using Lattice Boltzmann Techniques, and development of new algorithm to study nucleation and thermal flows with mesoscopic approaches are also under investigation. Among the most important impact for the society we cite the possibility to better assimilate data for numerical weather forecast, develop augmented reality approaches for bio-fluids and medicine, developing of optimal strategies for odor and contaminant sources in the presence of strong flow fluctuations. The overall objectives are to make a quantitative jump to data-driven approaches for fluid dynamics, offering open data suites to help the community for developing benchmarks and grand challenges,

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

We have developed:
(i) numerical algorithms for direct numerical simulations HPC applications of odor sources emissions in turbulent flows with mean wind.
(ii) developed Lattice Boltzmann Approaches for Nudging in Rayleigh Benard
(iii) Algorithm using Reinforcement Learning and Optimal Control equations for single and multiple Lagrangian probes in complex flows
(iv) algorithms using Generative Adversarial Networks (GAN) for Data Assimilation of geophysical flows
(v) algorithms using POD and EPOD to benchmark GAN approaches
(vi) Algorithms based on Lattice Boltzmann for multi-component micro and nano fluids
(viii) Algorithm based on Generative Diffusion Models for generation and data-assimilation of Lagrangian and Eulerian turbulent flows

Main results:
(i) first application of Nudging to Rayleigh Benard at high turbulent intensity
(ii) first application of GAN to reconstruct fully developed turbulent rotating flows
(iii) first application of adversarial reinforcement learning to study multi-agent prey-predator dynamics at low Reynolds
(iv) development of numerical tools for nano-particles using Lattice Boltzmann algorithms
(v) first application of Diffusion Models to generate complex Lagrangian turbulent trajectories
(vi) first application of Perseus and Deep-RL algorithm to solve for the Bellman optimality equation in Partial Observable Markov Decision Process for tracking of odors from fixed source in turbulence.
(vii) First application of Multi Objectives Reinforcement Learning for optimization of Lagrangian smart particles navigation in turbulent and stochastic flows
(viii) First application of Conditioned Diffusion Models to reconstruct Oceanic drifter's trajectories from partial measurements
(ix) First application of Nudging and Kalmann filters approach to reconstruct high and low frequencies fluctuations in turbulent models
(x) Investigation of smart Lagrangian forcing acting on concentrated bubbles in turbulent flows with 4-ways coupling regimes
(xi) Development of optimal sub-grid modeling using Machine Learning in the loop for turbulence
(xii) Application of generative AI tools to model charged particles trajectories for astrophysical problems, inertial particles for multi-phase turbulence
(xiii) Development of generative AI tools, based on Diffusion Models for modelling and reconstructing two-particle turbulent dispersion
(xiv) Development of quasi-optimal Reinforcement Learning protocols for odor search with egocentric measurements only in realistic 2d and 3d turbulent flows
(xv) Combination of Nudging plus generative AI tools (DM) for super-resolution of Lagrangian and Eulerian turbulence

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

We define a set of benchmarks and validation steps for the community working on Lagrangian and Eulerian data-driven tools for complex flows. Improving data assimilation, data augmentation, data generation abilities in many geophysical and applied situations where complex turbulent flows are acting. We expect to develop Physics-Informed Data-Driven new algorithms and applications to turbulent flows, and to develop novel forcing mechanisms to control turbulent transitions and properties. We deliver a portal for open-data to define benchmarks and grand challenges for the data-driven community working on complex fluids and complex flows properties.

screenshot-2023-12-06-at-13-54-01.png