Contact: Marcello MELDI.
The development and characterization of tools for the analysis of turbulent flows is the key subject of investigation of this transverse topic. The methodologies developed are employed for a large spectrum of applications which include flow control and data augmentation.
Within the framework of Industry 4.0, scientific activities described by keywords such as data augmentation and digital twins have seen rapid development in recent years for almost every branch of science. This is particularly true for fluid mechanics, where the increasing computational resources available at computational centers open new research perspectives in terms of data production and analysis. These include development and application of data-driven approaches relying on Uncertainty Quantification (UQ) and Data Assimilation (DA), as well as a second renaissance of approaches based on Machine Learning. Within the perimeter of LMFL and its scientific features, the Data Analysis transverse group federates research activities dealing with the following aspects:
- Data-driven analysis of instantaneous features and statistical evolution of high Reynolds regimes
- Model generation and reduction for flight dynamics
- State reconstruction and augmentation for complex flows
- Statistical inference for model optimization using frugal data-driven tools
- Online flow estimation and control
- Study of the attitude and the behaviour of an aircraft in the post-stall flight domain
List of recent and ongoing research projects in the laboratory
1. ANR-JCJC-2021-IWP-IBM-DA (2021-2025)
The accurate prediction of numerous flow features of unstationary flows such as aerodynamic forces is driven by the precise representation of localized near-wall dynamics. This aspect is particularly challenging for the flow prediction around complex geometries. In this case classical body-fitted approaches may have to deal with high deformation of the mesh elements, possibly providing poor numerical prediction. Additionally, the simulation of moving bodies may require prohibitively expensive mesh updates. The Immersed Boundary Method (IBM) has emerged as one of the most popular strategies to handle these two problematic aspects. The main difficulty of IBM is the high computational demands for the representation of wall turbulence, which is a governing aspect in most engineering cases.
Project IWP-IBM-DA aims to obtain advancement of the IBM method via DA. The accuracy of the IBM will be improved by integrating high-fidelity observation (including experimental data), targeting a precise representation of near-wall dynamics for turbulent flows. The bullet points of the present proposal representing innovation with respect to the state-of-the-art are i) the research development is conceived for online DA applications, ii) DA is coupled with Machine Learning (ML), which provides new predictive tools not relying on observation and iii) the application to complex, realistic cases is envisioned.
2. ONERA intern research project – Measurement of the danger of a spin by
Loss of Control in flight (LOC-I) accidents remain the major cause of fatal aircraft’s accidents, for both general and commercial aviation. Several factors, such as the dysfunction of instrumentation, incorrect control inputs or sudden flow events, may lead to an unwanted excursion in the post-stall flight domain which can induce LOC-I accidents. In flight dynamics, the dynamical system theory is used to study the different behaviours observed in the post-stall domain. Indeed, dangerous behaviours, as spins or deep-stall, can be linked to a stable steady-state of the dynamical system composed of the differential equations traducing the motion of the aircraft. Since the 1980’s, the bifurcation theory has been extensively used to have a global overview of the possible behaviours encountered in the post-stall domain. One of the remaining challenges is to estimate the region of attraction of the steady state, representing the set of initial conditions leading to the steady-state. The knowledge of this region of attraction improves the quality of the information on the predicted behaviour in the post-stall domain. The bigger the region of attraction is in the phase space, the more danger the behaviour represents, as the set of initial conditions leading to that behaviour occupies a large space (in terms of phase space).
The aim of this ONERA intern research project is the use of machine learning to estimate the regions of attraction of dangerous post-stall equilibrium points. New modern methods of estimation of region of attraction, based on Gaussian Process classifiers (GPc) or Physics Informed Neural Networks (PINN) are tested on predicted equilibrium points obtained with an aircraft model. The performances of the new methods will be compared with the ones of the classical method relying on Sum of Squares Polynomials (SOS), which is built on strong simplification assumptions.
The three figures below compare the results of the methods (respectively SOS, GPc and PINN) of estimation of the region of attraction of the time-inverted Van der Pol oscillator system, a classical dynamical system with 2 state variables which is used to validate the implementation of the methods.
3. ANR-JCJC-2021-MultiMatchGrid (2021-2025)
Laminar and transitional flows strongly perturbed by the presence of a weakly-porous rigid grid will be considered. A wide spectrum of scales is involved in such flows, ranging from the largest convective scales of the incoming flow down to the small grid-hole-diameters scale. Future researchers will develop a first-principle approach to deal with weakly-porous grids by exploiting the separation of scales between the large-scale flow far from the grid and the small-scale flow at the grid’s pores’ scale. A novel and efficient scale-matching approach will be derived by making use of theoretical analysis, numerical simulations, and recent machine learning advancements (see figure below).
Novel multi-scale first-principle criteria will be derived within the framework of the ANR-JCJC project MultiMatchGrid in order to model and carry out predictive simulations of grid flows. The long-term goal of such an approach is improving the robustness and security of grid-based flow controls and avoiding undesired flow instabilities. The resulting methodology will be spendable by the industry for designing grid-based passive control systems that will allow an efficient and robust design of potentially ground-breaking mixing enhancers. Other potential extensions include turbulence control in pipes and inflow control for turbomachinery. Future generalizations may include the extension of our methodology to deformable and moving grids, with potential applications to typical textile design problems and complex membrane dynamics.