Gianluca Iaccarino

In this work, an LES model with finite-rate phase change and subgrid spray based on a high-resolution numerical scheme for multiphase multi-component simulations which satisfies interface equilibrium and phase immiscibility conditions is proposed. The multiphase model is based on a robust implementation of the four-equation multiphase model which assumes a strict subgrid equilibrium of pressure, temperature, and velocity. Critically, the equilibrium assumptions of the four-equation model provide large computational savings compared to modeling the full non-equilibrium multiphase system. To obtain predictive capabilities with these restrictive equilibrium assumptions, a new phase-confined form of the Eulerian $Σ$ spray model is proposed to predict subgrid interfacial surface area while avoiding unphysical leakage across interfaces. Additionally, an improved finite rate phase change model which is thermodynamically bounded by the equilibration of the Gibbs-free energy is coupled with the $Σ$ equation to model complex phase change regimes. The full modeling framework is validated using the Engine Combustion Network (ECN) Spray A case in non-evaporating and evaporating conditions and shows excellent agreement with experimental measurements.

In this work, an LES model with finite-rate phase change and subgrid spray based on a high-resolution numerical scheme for multiphase multi-component simulations which satisfies interface equilibrium and phase immiscibility conditions is proposed. The multiphase model is based on a robust implementation of the four-equation multiphase model which assumes a strict subgrid equilibrium of pressure, temperature, and velocity. Critically, the equilibrium assumptions of the four-equation model provide large computational savings compared to modeling the full non-equilibrium multiphase system. To obtain predictive capabilities with these restrictive equilibrium assumptions, a new phase-confined form of the Eulerian $Σ$ spray model is proposed to predict subgrid interfacial surface area while avoiding unphysical leakage across interfaces. Additionally, an improved finite rate phase change model which is thermodynamically bounded by the equilibration of the Gibbs-free energy is coupled with the $Σ$ equation to model complex phase change regimes. The full modeling framework is validated using the Engine Combustion Network (ECN) Spray A case in non-evaporating and evaporating conditions and shows excellent agreement with experimental measurements.

Accurate and predictive scale-resolving simulations of laser-ignited rocket engines are highly time-consuming because the problem includes turbulent fuel-oxidizer mixing dynamics, laser-induced energy deposition, and high-speed flame growth. This is conflated with the large design space primarily corresponding to the laser operating conditions and target location. To enable rapid exploration and uncertainty quantification, we propose a data-driven surrogate modeling approach that combines convolutional autoencoders (cAEs) with neural ordinary differential equations (neural ODEs). The present target application of an ML-based surrogate model to leading-edge multi-physics turbulence simulation is part of a paradigm shift in the deployment of surrogate models towards increasing real-world complexity. Sequentially, the cAE spatially compresses high-dimensional flow fields into a low-dimensional latent space, wherein the system's temporal dynamics are learned via neural ODEs. Once trained, the model generates fast spatiotemporal predictions from initial conditions and specified operating inputs. By learning a surrogate to replace the entirety of the time-evolving simulation, the cost of predicting an ignition trial is reduced by several orders of magnitude, allowing efficient exploration of the input parameter space. Further, as the current framework yields a spatiotemporal field prediction, appraisal of the model output's physical grounding is more tractable. This approach marks a significant step toward real-time digital twins for laser-ignited rocket combustors and represents surrogate modeling in a complex system context.

Accurate and predictive scale-resolving simulations of laser-ignited rocket engines are highly time-consuming because the problem includes turbulent fuel-oxidizer mixing dynamics, laser-induced energy deposition, and high-speed flame growth. This is conflated with the large design space primarily corresponding to the laser operating conditions and target location. To enable rapid exploration and uncertainty quantification, we propose a data-driven surrogate modeling approach that combines convolutional autoencoders (cAEs) with neural ordinary differential equations (neural ODEs). The present target application of an ML-based surrogate model to leading-edge multi-physics turbulence simulation is part of a paradigm shift in the deployment of surrogate models towards increasing real-world complexity. Sequentially, the cAE spatially compresses high-dimensional flow fields into a low-dimensional latent space, wherein the system's temporal dynamics are learned via neural ODEs. Once trained, the model generates fast spatiotemporal predictions from initial conditions and specified operating inputs. By learning a surrogate to replace the entirety of the time-evolving simulation, the cost of predicting an ignition trial is reduced by several orders of magnitude, allowing efficient exploration of the input parameter space. Further, as the current framework yields a spatiotemporal field prediction, appraisal of the model output's physical grounding is more tractable. This approach marks a significant step toward real-time digital twins for laser-ignited rocket combustors and represents surrogate modeling in a complex system context.

To understand the reliability of successful ignition through use of laser energy deposition for a liquid-oxygen gaseous-methane rocket combustor, we conduct large-eddy simulations (LESs). We first ensure that pre-ignition flow statistics accurately reproduce companion experimental measurements conducted at Purdue University. With satisfactory agreement observed in pre-ignition, we vary physical input parameters to construct an ensemble of LES ignition trials. The majority of the computations are coarse LES but a small set of enhanced refinement calculations with an order of magnitude more grid points verify that the ignition mechanism is the same across fidelities. Subsequently, alignment between experiments and simulations is enabled by a parameter estimation framework, whereby unmeasurable quantities associated with the laser operation are inferred by the uncertainty in underlying successful ignition likelihoods that arises from a limited number of experimental trials. With confidence established in the LES, we subsequently report the ignition pathways contrasting the critical instances that lead to ignition failure or success.

Robots with multi-modal locomotion are an active research field due to their versatility in diverse environments. In this context, additional actuation can provide humanoid robots with aerial capabilities. Flying humanoid robots face challenges in modeling and control, particularly with aerodynamic forces. This paper addresses these challenges from a technological and scientific standpoint. The technological contribution includes the mechanical design of iRonCub-Mk1, a jet-powered humanoid robot, optimized for jet engine integration, and hardware modifications for wind tunnel experiments on humanoid robots for precise aerodynamic forces and surface pressure measurements. The scientific contribution offers a comprehensive approach to model and control aerodynamic forces using classical and learning techniques. Computational Fluid Dynamics (CFD) simulations calculate aerodynamic forces, validated through wind tunnel experiments on iRonCub-Mk1. An automated CFD framework expands the aerodynamic dataset, enabling the training of a Deep Neural Network and a linear regression model. These models are integrated into a simulator for designing aerodynamic-aware controllers, validated through flight simulations and balancing experiments on the iRonCub-Mk1 physical prototype. Flying humanoid robots face challenges in modelling and control, particularly with aerodynamic forces. Antonello Paolino and colleagues propose the mechanical design of iRonCub-Mk1, a jet-powered humanoid robot, and a methodology to estimate, validate, and control aerodynamic forces during flight.

This study numerically investigates two sources of uncertainties that may influence measurements of flow past a circular cylinder confined in a duct. These being spanwise asymmetries in upstream profiles and measurement plane location. We find variations in upstream profiles strongly affects wake topology. Whereas, due to end effects, uncertainties in measurement plane location strongly influences measurement of flow statistics. Hence, a combination of both may yield uncertainties in flow measurements. These insights may be used to guide experimental investigation of moderate to highly confined flows.

We investigate methods for mapping between low- and high-resolution simulations to generate surrogate models, which would significantly reduce the overall computational costs. Our focus is on interpolative decomposition, a rank-revealing matrix decomposition technique that efficiently selects key parameters to minimize the number of required simulation sets. We demonstrate that this approach remains effective even in the presence of multiple flow regimes (mode transitions) and show that the mapping process can also function as a classification tool for identifying different flow modes.

A novel strategy for generating datasets has been developed within the context of drag prediction for automotive geometries using neural networks. A primary challenge in this field is constructing a training database of sufficient size and diversity. Our method relies on a small number of initial data points and provides a recipe to systematically interpolate between them, generating an arbitrary number of samples at the desired quality. We tested this strategy using a representative automotive geometry and demonstrated that convolutional neural networks perform exceptionally well at predicting drag coefficients and surface pressures. Promising results were obtained in testing extrapolation performance. Our method can be applied to other problems of aerodynamic shape optimization.

To predict the reliability of laser ignition in a rocket combustor using large-eddy simulations (LESs), it is essential to first ensure that the pre-ignition jet statistics and the dynamics of the hot kernel generated by the energy deposition are accurately captured. In this manuscript, we compare numerical results with experimental data to evaluate the accuracy of the computational approach. First, the jet LES statistics show good qualitative agreement with the particle imaging velocimetry (PIV) data. Quantitative comparisons at several streamwise locations reveal larger differences near the injector, but with local discrepancies of less than 15 m/s in both the mean and fluctuation statistics. Second, we quantify the mean and uncertainties of the hot kernel modeling parameters through a joint analysis of experimental data and direct numerical simulation (DNS) results. This approach accounts for shot-to-shot variability in the simulations, which demonstrate good agreement with the experimental data regarding the ejecta position.

Laser-induced spark is an ignition method in a rocket combustor that facilitates engine re-ignition throughout a mission. The present work generates a probability map of successful ignition that varies with laser deposition site, and also, a quantitative framework for the post-ignition pressure rise probability distribution in the combustor. Run-to-run variabilities in the system arise from uncertainty in the deposited laser energy kernel characteristics, and the stochastic nature of turbulence. Multi-fidelity Monte Carlo sampling is implemented to obtain realizations of the uncertainty space. Low fidelity is achieved by coarsening of the mesh used for simulation, and simplification of the underlying chemistry. We demonstrate the application of a novel bi-fidelity method, the \textit{stochastic interpolative decomposition}, which enables the estimation of the quantity of interest for the equivalent high fidelity ensemble using a small number of runs. Bi-fidelity stochastic interpolative decomposition is a suitable uncertainty quantification tool for problems exhibiting stochastic, multi-modal outcomes in the uncertainty space, as it preserves correlations in the output space where standard interpolative decomposition fails.

In aerodynamic design, the use of Computational Fluid Dynamics (CFD) tools has long relied on Reynolds-Averaged Navier-Stokes (RANS) and Unsteady RANS (URANS) turbulence models. However, these methods face limitations in predicting massively separated flow regions. As a result, there is growing emphasis on scale-resolving methods such as Large Eddy Simulation (LES), which have become more feasible due to advancements in computational resources. This study focuses on the LES of airfoil flows, specifically analyzing the influence of the spanwise extent of the computational domain on aerodynamic forces. We examine the NREL S826 airfoil at a chord Reynolds number of $Re_c = 10^5$ under different angles of attack, with simulations performed using the PyFR solver. Our results highlight the importance of adequate spanwise resolution, particularly in post-stall flow conditions, where larger span sizes are necessary to accurately capture three-dimensional flow structures. The findings suggest that for certain flow regimes, traditional two-dimensional statistical representations may not be sufficient, necessitating further investigation into 3D flow phenomena.

Data from direct numerical simulations of turbulent flows are commonly used to train neural network-based models as subgrid closures for large-eddy simulations; however, models with low a priori accuracy have been observed to fortuitously provide better a posteriori results than models with high a priori accuracy. This anomaly can be traced to a dataset shift in the learning problem, arising from inconsistent filtering in the training and testing stages. We propose a resolution to this issue that uses explicit filtering of the nonlinear advection term in the large-eddy simulation momentum equations to control aliasing errors. Within the context of explicitly-filtered large-eddy simulations, we develop neural network-based models for which a priori accuracy is a good predictor of a posteriori performance. We evaluate the proposed method in a large-eddy simulation of a turbulent flow in a plane channel at $Re_τ = 180$. Our findings show that an explicitly-filtered large-eddy simulation with a filter-to-grid ratio of 2 sufficiently controls the numerical errors so as to allow for accurate and stable simulations.

A novel strategy for generating datasets is developed within the context of drag prediction for automotive geometries using neural networks. A primary challenge in this space is constructing a training databse of sufficient size and diversity. Our method relies on a small number of starting data points, and provides a recipe to interpolate systematically between them, generating an arbitrary number of samples at the desired quality. We test this strategy using a realistic automotive geometry, and demonstrate that convolutional neural networks perform exceedingly well at predicting drag coefficients and surface pressures. Promising results are obtained in testing extrapolation performance. Our method can be applied to other problems of aerodynamic shape optimization.

A data-driven model is compared to classical equation-driven approaches to investigate its ability to predict quantity of interest and their uncertainty when studying airfoil aerodynamics. The focus is on autoencoders and the effect of uncertainties due to the architecture, the hyperparamaters and the choice of the training data (internal or model-form uncertainties). Comparisons with a Gaussian Process regression approach clearly illustrate the autoencoder advantage in extracting useful information on the prediction confidence even in the absence of ground truth data. Simulations accounting for internal uncertainties are also compared to the impact of the variability induced by uncertain operating conditions (external uncertainties) showing the importance of accounting for the total uncertainty when establishing prediction confidence.

BACKGROUND: Impaired cerebrospinal fluid (CSF) dynamics is involved in the pathophysiology of neurodegenerative diseases of the central nervous system and the optic nerve (ON), including Alzheimer's and Parkinson's disease, as well as frontotemporal dementia. The smallness and intricate architecture of the optic nerve subarachnoid space (ONSAS) hamper accurate measurements of CSF dynamics in this space, and effects of geometrical changes due to pathophysiological processes remain unclear. The aim of this study is to investigate CSF dynamics and its response to structural alterations of the ONSAS, from first principles, with supercomputers. METHODS: Large-scale in-silico investigations were performed by means of computational fluid dynamics (CFD) analysis. High-order direct numerical simulations (DNS) have been carried out on ONSAS geometry at a resolution of 1.625 μm/pixel. Morphological changes on the ONSAS microstructure have been examined in relation to CSF pressure gradient (CSFPG) and wall strain rate, a quantitative proxy for mass transfer of solutes. RESULTS: A physiological flow speed of 0.5 mm/s is achieved by imposing a hydrostatic pressure gradient of 0.37-0.67 Pa/mm across the ONSAS structure. At constant volumetric rate, the relationship between pressure gradient and CSF-accessible volume is well captured by an exponential curve. The ONSAS microstructure exhibits superior mass transfer compared to other geometrical shapes considered. An ONSAS featuring no microstructure displays a threefold smaller surface area, and a 17-fold decrease in mass transfer rate. Moreover, ONSAS trabeculae seem key players in mass transfer. CONCLUSIONS: The present analysis suggests that a pressure drop of 0.1-0.2 mmHg over 4 cm is sufficient to steadily drive CSF through the entire subarachnoid space. Despite low hydraulic resistance, great heterogeneity in flow speeds puts certain areas of the ONSAS at risk of stagnation. Alterations of the ONSAS architecture aimed at mimicking pathological conditions highlight direct relationships between CSF volume and drainage capability. Compared to the morphological manipulations considered herein, the original ONSAS architecture seems optimized towards providing maximum mass transfer across a wide range of pressure gradients and volumetric rates, with emphasis on trabecular structures. This might shed light on pathophysiological processes leading to damage associated with insufficient CSF flow in patients with optic nerve compartment syndrome.

Inverse design, where we seek to design input variables in order to optimize an underlying objective function, is an important problem that arises across fields such as mechanical engineering to aerospace engineering. Inverse design is typically formulated as an optimization problem, with recent works leveraging optimization across learned dynamics models. However, as models are optimized they tend to fall into adversarial modes, preventing effective sampling. We illustrate that by instead optimizing over the learned energy function captured by the diffusion model, we can avoid such adversarial examples and significantly improve design performance. We further illustrate how such a design system is compositional, enabling us to combine multiple different diffusion models representing subcomponents of our desired system to design systems with every specified component. In an N-body interaction task and a challenging 2D multi-airfoil design task, we demonstrate that by composing the learned diffusion model at test time, our method allows us to design initial states and boundary shapes that are more complex than those in the training data. Our method generalizes to more objects for N-body dataset and discovers formation flying to minimize drag in the multi-airfoil design task. Project website and code can be found at https://github.com/AI4Science-WestlakeU/cindm.

.In recent years there has been a push to discover the governing equations of dynamical systems directly from measurements of the state, often motivated by systems that are too complex to directly model. Although there has been substantial work put into such a discovery, doing so in the case of large noise has proved challenging. Here we develop an algorithm for the simultaneous identification and denoising of a dynamical system (SIDDS). We infer the noise in the state measurements by requiring that the denoised state satisfies the dynamical system with an equality constraint. This contrasts to existing work where the mismatch in the dynamics is added as a penalty in the objective. Assuming the nonlinear differential equation is represented in a predefined basis, we develop a sequential quadratic programming approach to solve the SIDDS problem featuring a direct solution of the KKT system with a specialized preconditioner. We also show how to add a sparsity promotion regularization into SIDDS using an iteratively reweighted least squares approach. Our resulting algorithm obtains estimates of the dynamical system that achieve the Cramér–Rao lower bound up to discretization error. This enables SIDDS to provide substantial improvements compared to existing techniques: SIDDS substantially decreases the data burden for accurate identification, recovers optimal estimates with lower sample rates, and the sparsity promoting variant discovers the correct sparsity pattern with larger noise.Keywordsdynamical systemsmodel discoveryinverse problemsparameter estimationsparse recoveryMSC codes34A5565L0990C5593B30

View Video Presentation: https://doi.org/10.2514/6.2023-4364.vid An autoencoder is trained to reproduce the aerodynamic characteristics of wing sections (2D airfoils). The formulation is based on an adaptive training database that minimize the data required while preserving the accuracy of the solutions. The autoencoder uses aggressive compression (small latent dimension) to mimic the independent variables used to define the database. The latent space is interpolated using Radial Basis Functions with a cubic kernel to generate synthetic flow fields on unseen airfoils. The accuracy of the results and the interpretation of the latent space are based on comparisons with simulations.

View Video Presentation: https://doi.org/10.2514/6.2023-3597.vid An integrated computational framework is introduced to study complex engineering systems through physics-based ensemble simulations on heterogeneous supercomputers. The framework is primarily designed for the quantitative assessment of laser-induced ignition in rocket engines. We develop and combine an implicit programming system, a compressible reacting flow solver, and a data generation/management strategy on a robust and portable platform. We systematically present this framework using test problems on a hybrid CPU/GPU machine. Efficiency, scalability, and accuracy of the solver are comprehensively assessed with canonical unit problems. Ensemble data management and autoencoding are demonstrated using a canonical diffusion flame case. Sensitivity analysis of the ignition of a turbulent, gaseous fuel jet is performed using a simplified, three-dimensional model combustor. Our approach unifies computer science, physics and engineering, and data science to realize a cross-disciplinary workflow. The framework is exascale-oriented and can be considered a benchmark for future computational science studies of real-world systems.

Wake steering yaws upstream wind turbines to deflect their wakes from downstream turbines, increasing the total power produced by the wind farm. Most wake steering methods generate lookup tables offline which map a set of wind farm conditions, such as wind speed, to yaw offset angles for each turbine in a farm. These tables assume all turbines are operational and can be significantly non-optimal when one or more turbines shutdown–as they often do because of low wind speed, routine maintenance, or emergency maintenance. We present a new wake steering method that adapts to turbine status. Using a hybrid model- and learning-based method, differentiable control, we train a neural network to determine yaw offset angles from conditions including turbine status (active/inactive). Unlike the lookup table approach, differentiable control does not solve an optimization problem for each combination of turbine status in a farm; including learning in the method allows it to generalize. We present results for both standard wake steering (all turbines active) and adaptive wake steering (some turbines active). We find that differentiable control has comparable accuracy as and an order of magnitude faster offline compute time than the lookup table approach. Differentiable control enables adaptive wake steering through computationally efficient training and rapid online evaluation.

Applications of machine learning techniques to physics applications are gaining popularity by providing solution process acceleration, super-resolution, equation discovery, data compression, etc. Different flavors of data driven techniques have been proposed in the literature, in the first part of this talk, I will focus on hybrid approaches in which an existing grid-based numerical technique for solving fluid governing equations is augmented with a neural network correction. It is unclear if these approaches provide improvements over the baseline numerical solver by approximating local subgrid-scale dynamics, by incorporating long-range spatial/temporal correlations or by introducing high-order-like solution reconstructions. Furthermore, it is unclear how different aspects of the training pipeline affect the quality of the neural correction and its generalizability. We investigate the issues above in turbulent channel flow and thermal convection problems. In the second part of the talk, I will focus on the use of convolutional autoencoders to study aerodynamic stall and focus on extreme latent space compression to extract physical insights. We show how physics imprinting (the correlation between latent variable and physical quantities) naturally emerges and is affected by the training database.