Deep learning is rapidly emerging as a powerful tool for surrogate modeling and control in Computational Fluid Dynamics (CFD). Conceptually, these data-driven approaches are reshaping scientific practice, reviving the longstanding debate between data-driven and equation-based modeling. A central question is whether, and how, such methods can challenge or complement traditional CFD, which relies on well-established mathematical formulations. In this seminar, I will first outline the key principles that underpin efficient and reliable scientific modeling. I will then discuss how modern AI techniques can be designed to incorporate these principles and achieve practical effectiveness. In particular, I will highlight the role of implicit neural representations (INR) and uncertainty quantification (UQ) as critical components for robust and accurate surrogate modeling in aerospace engineering.
Catégorie d'événements: Fluids, Thermal and Combustion
A statistical theory of disturbance growth in transitional flows
The growth of small disturbances and the subsequent transition from laminar to turbulent flow is of central importance in many engineering applications, including high-speed flight, inertial confinement fusion, and noise pollution from airplanes and wind turbines. The growth of disturbances in fluids has traditionally been characterized by their long-time asymptotic stability and, more recently, by the optimal transient (finite-time) amplification maximized over all possible initial disturbances. These descriptions, while valuable, are limited in numerous ways: the asymptotic growth or decay may emerge only on unreasonably long timescales, real disturbances typically grow much less than the finite-time optimal disturbance, and both methods rely on linearized governing equations that are not always accessible. We have developed a suite of statistical and data-driven tools to overcome these limitations and provide a more comprehensive description of the growth of disturbances in fluid systems. Given a statistical description of the initial disturbances, our framework predicts the expected (mean) growth, a hierarchy of coherent structures responsible for the expected growth, and the probability of observing any particular level of amplification, i.e., the probability density function of the evolved disturbances. Additionally, we have developed a data-driven implementation of the optimal transient growth theory that approximates the upper bound on the growth of any initial disturbance given finite data. We are currently using these new tools to develop a probabilistic transition model for hypersonic boundary layers and to assess the impact of surface imperfections of fuel capsules for inertial confinement fusion.
Rapid, Accurate, and Robust Predictions and Tracking for Space Domain Awareness
Space domain awareness (SDA) comprises acquiring observations and processing those observations to perform orbit determination (OD), correlation and catalogue updates, tracking, predictions, and conjunction analysis. The existing SDA infrastructure provided by the US Space Surveillance Network (SSN) and the European Space Agency (ESA) Space Surveillance and Tracking (SST) is at its limit due to the continuous growth of Earth orbiting constellations, and the projected exponential growth of orbital debris (Kessler’s syndrome). In this talk, we will present our research developing the scientific framework to achieve rapid uncertainty quantification and propagation for space domain awareness and decision making. This enables domain awareness operations on the ground as well as onboard in real-time to resolve detected objects and predict their future locations. The present rapid uncertainty quantification combines model-based approximation techniques and data-driven approaches to quantify detected/tracked objects uncertainties in real-time. Results of the proposed model-driven uncertainty propagation approach have shown 2 – 3 orders of magnitude improvement over existing methods. The data-driven approach trained on the high-fidelity model is aimed to be implemented on spaceflight hardware. Highlighted in the present algorithm are the outcomes of the collaboration with professor Razaaly via the Chateubriand Fellowship.
Tracking the nonlinear formation of an interfacial wave cascade: from one, to few, to many
A hallmark of far-from-equilibrium physics is the emergence of a spectral cascade, where energy is transferred across length-scales following a simple power law. Scaling laws of steady states have been successfully predicted in fluid dynamics by the statistical theory of weak wave turbulence. However, the predictive power of this theory breaks down in presence of large amplitudes, high dissipation, and finite size effects. We offer experimental insight into these regimes by resolving the dynamics of individual wave modes in an externally driven fluid-fluid interface. We observe the time evolution of interface excitations from one to few to many, a process culminating in a direct energy cascade. Our findings confirm that interfacial dynamics can be effectively modelled by a weakly nonlinear Lagrangian theory, revealing a hierarchy in their order that confirms a key assumption of weak wave turbulence. Specific interactions are tracked through time, and we predict the timescale until a cascade emerges. Furthermore, we highlight how our experiment may inform us about other far-from-equilibrium systems by mapping our Lagrangian theory to a model of cosmological preheating in the early universe.
Reinforcement Twinning and the Reciprocal Learning of Models and Control Policies
Optimal control and reinforcement learning are often viewed as competing paradigms for sequential decision-making. While model-based control relies on adjoint sensitivities, reinforcement learning derives policies through value-function approximation. However, the Hamiltonian in optimal control and the state–action value function in reinforcement learning plays analogous roles in guiding policy improvement.
This work leverages these structural correspondences to develop mechanisms of reciprocal reinforcement between modelling and control within the Reinforcement Twinning framework. In this cyber–physical architecture, a digital twin and a learning agent evolve within a shared feedback loop: the twin provides the predictive structure and sensitivity information necessary to accelerate policy optimization, while the agent generates informative trajectories that improve the model.
Illustrative test cases on nonlinear thermo-fluid systems demonstrate how policies optimized on imperfect twins can extract robust operational knowledge under stochastic disturbances. By explicitly exploiting and refining model uncertainty rather than assuming absolute fidelity, this framework demonstrates that the synergy between predictive modelling and autonomous learning outperforms their independent implementation, providing a more resilient path for the control of complex, nonlinear systems.
Genius in shallow water ship hydrodynamics
Shallow water ship hydrodynamics is a highly diverse field riddled with paradoxical individuals in its history. Where they were able to control the narrative of their stories, fame followed, and conversely, where they were unable to do so, they languished in obscurity. This talk will briefly explore the lives of two such diametrically opposed individuals, showing that although genius can come in many forms, controlling one’s narrative is what makes or breaks legacy.
Modelling jet-plate interaction with a resonant normal form
After reviewing the experimental evidence for screech in a supersonic jet with and without a thin plate, we seek a minimal dynamical description capable of reproducing the observed changes in staging and spectral content. The experiments show that, in the unperturbed axisymmetric configuration, the classical sequence features an abrupt transition from mode B (flapping, standing-wave–like) to mode C (spinning, rotating-wave–like). When a plate is inserted parallel to the jet axis, this jump is suppressed and replaced by a continuous mixed response; in addition, in selected Mach number ranges the PSD exhibits two nearby spectral lines consistent with time-dependent mixed states. Motivated by these observations, we start from the resonant 1:1 Hopf normal form governing the azimuthal pair |m|=1associated with modes B and C, and introduce a symmetry-breaking perturbation representing the jet–plate interaction. The plate destroys continuous rotational invariance while preserving a reflection symmetry, thereby enabling coupling terms that are forbidden in the O(2)-equivariant case. We then classify the resulting solution branches (pure even/odd states and mixed states) and distinguish between phase-locked solutions, producing a single dominant tone, and phase-drifting (unlocked) solutions, producing a spectral line splitting through an Adler-type phase-drift mechanism. By comparing the qualitative bifurcation structure and predicted spectral signatures with the experimental PSDs, we show how the perturbed normal form naturally explains both the disappearance of the B→C jump and the emergence of mixed states. Finally, we highlight a possible amplitude-symmetry–breaking (pitchfork) bifurcation within the mixed family, providing a dynamical route to unequal-amplitude mixed states consistent with the asymmetric spectral weights measured by symmetric microphones
High-Compressible flows – Low frequency unsteadiness and Tonal Mechanisms
In this presentation, we will examine how the wake generated by micro-vortex generators or the presence of an upstream sweep affects the physics of a canonical Shock–Boundary Layer Interactions (SBLI) using high-fidelity numerical simulations and advanced post-processing techniques to better understand the underlying flow dynamics.
We will then explore how, in many compressible flows, different types of waves—aerodynamic, acoustic, and entropic—interact and couple, sometimes generating strong oscillations and pronounced tonal noise at specific frequencies. These waves have distinct characteristic velocities and typically interact at flow singularities, such as walls, corners, or shocks.
Several configurations will be discussed, including non-ideally expanded jets issuing from pipe nozzles and TIC nozzles, as well as supersonic air intakes. All of these systems exhibit significant flow oscillations driven by wave coupling.
Finally, we will look at the low-Mach-number regime (M < 0.1), where most feedback mechanisms disappear. Interestingly, the aeroacoustic feedback loop responsible for airfoil tonal noise remains active.
Why does this mechanism persist when others vanish?
Modelling and control of coherent turbulent structures in aerodynamics and aeroacoustics
Despite the intrinsic complexity of turbulent flows, large-scale coherent structures are present in turbulence, and are related to aerodynamic drag and sound radiation, for example. This presentation shows methods to model such coherent structures, and approaches to control flows in order to reduce drag and noise radiation. Linear models will be considered first, with coherent structures modelled as dominant responses from the linearised Navier-Stokes system. Such dominant modes compare favourably with coherent structures educed from experimental or numerical data, and allow the proposition of changes to the system aiming at drag or noise reduction; moreover, such linearised models allow the formulation of estimation and control problems. Non-linear reduced-order models (ROMs) are then obtained by the Galerkin projection of the full governing equations onto a basis of modes obtained from the linearised system. Quantitative agreement with reference statistics is obtained for a number of canonical flow configurations. Finally, a ROM for turbulent Couette flow is used to devise a control method aiming at turbulence suppresion. The obtained strategy is seen to relaminarise turbulence in direct numerical simulations, showing that ROMs so obtained have promising applications in optimisation and control.
Wake-Induced Laminar-Turbulent Transition in Separated Boundary Layers.
Wake-induced laminar-to-turbulent bypass transition in a separated laminar boundary layer (SLB) is investigated downstream of a cylinder of diameter D mounted in a constant free stream near a smooth flat plate. Direct numerical simulation (DNS) is conducted for a moderate gap of 0.9D with _ = 3900 and _ = 150, based on the momentum thickness. The transition process is driven by coherent vortex dynamics along the wall synchronized with Kármán shedding behind the cylinder. The transition process is identified from the mean velocity and Reynolds stress fields and then characterized from the spatial evolution o the velocity field spectra. Turbulence emerges at such remarkably low _ because the transition process follows a hybrid pathway combining SLB instability with wake-driven near-wall Λ-vortex formation and their interaction with periodically shed vortices. The transition unfolds in distinct stages: linear disturbance amplification; nonlinear saturation via super-harmonic resonance; and turbulent breakdown. Unlike classical SLB transitions, the wake’s periodicity imposes unique spectral signatures on the transition dynamics. Based on weakly nonlinear theory, an approach for analyzing the nonlinear interactions and spectral energy transfer between scales is proposed. The findings characterize wake-boundary interactions inherent to turbomachinery cascades or slotted-wing aerodynamic systems, offering insights for potential flow control strategies.