Search

This work develops a distributed graph neural network (GNN) methodology for mesh-based modeling applications using a consistent neural message passing layer. As the name implies, the focus is on enabling scalable operations that satisfy physical consistency via halo nodes at sub-graph boundaries. Here, consistency refers to the fact that a GNN trained and evaluated on one rank (one large graph) is arithmetically equivalent to evaluations on multiple ranks (a partitioned graph). This concept is demonstrated by interfacing GNNs with NekRS, a GPU-capable exascale CFD solver developed at Argonne National Laboratory. It is shown how the NekRS mesh partitioning can be linked to the distributed GNN training and inference routines, resulting in a scalable mesh-based data-driven modeling workflow. We study the impact of consistency on the scalability of mesh-based GNNs, demonstrating efficient scaling in consistent GNNs for up to O(1B) graph nodes on the Frontier exascale supercomputer.

A graph neural network (GNN) approach is introduced in this work which enables mesh-based three-dimensional super-resolution of fluid flows. In this framework, the GNN is designed to operate not on the full mesh-based field at once, but on localized meshes of elements (or cells) directly. To facilitate mesh-based GNN representations in a manner similar to spectral (or finite) element discretizations, a baseline GNN layer (termed a message passing layer, which updates local node properties) is modified to account for synchronization of coincident graph nodes, rendering compatibility with commonly used element-based mesh connectivities. The architecture is multiscale in nature, and is comprised of a combination of coarse-scale and fine-scale message passing layer sequences (termed processors) separated by a graph unpooling layer. The coarse-scale processor embeds a query element (alongside a set number of neighboring coarse elements) into a single latent graph representation using coarse-scale synchronized message passing over the element neighborhood, and the fine-scale processor leverages additional message passing operations on this latent graph to correct for interpolation errors. Demonstration studies are performed using hexahedral mesh-based data from Taylor-Green Vortex flow simulations at Reynolds numbers of 1600 and 3200. Through analysis of both global and local errors, the results ultimately show how the GNN is able to produce accurate super-resolved fields compared to targets in both coarse-scale and multiscale model configurations.

Numerical simulations of detonation-containing flows have emerged as crucial tools for designing next-generation power and propulsion devices. As these tools mature, it is important for the combustion community to properly understand and isolate grid resolution effects when simulating detonations. To this end, this work provides a comprehensive analysis of the numerical convergence of unsteady detonation simulations, with focus on isolating the impacts of chemical timescale modifications on convergence characteristics in the context of operator splitting. With the aid of an adaptive mesh refinement based flow solver, the convergence analysis is conducted using two kinetics configurations: (1) a simplified three-step model mechanism, in which chemical timescales in the detonation are modified by adjusting activation energies, and (2) a detailed hydrogen mechanism, in which chemical timescales are adjusted through ambient pressure modifications. The convergence of unsteady self-sustained detonations in one-dimensional channels is then analyzed with reference to steady-state theoretical baseline solutions using these mechanisms. The goal of the analysis is to provide a detailed comparison of the effects of grid resolution on both macroscopic (peak pressures and detonation wave speeds) and microscopic (detonation wave structure) quantities of interest, drawing connections between the deviations from steady-state baselines and minimum chemical timescales. This work uncovers resolution-dependent unsteady detonation regimes, and highlights the important role played by not only the chemical timescales, but also the ratio between chemical and induction timescales in the detonation wave structure on simulation convergence properties.

In this work, we provide a mathematical formulation for error propagation in flow trajectory prediction using data-driven turbulence closure modeling. Under the assumption that the predicted state of a large eddy simulation prediction must be close to that of a subsampled direct numerical simulation, we retrieve an upper bound for the prediction error when utilizing a data-driven closure model. We also demonstrate that this error is significantly affected by the time step size and the Jacobian which play a role in amplifying the initial one-step error made by using the closure. Our analysis also shows that the error propagates exponentially with rollout time and the upper bound of the system Jacobian which is itself influenced by the Jacobian of the closure formulation. These findings could enable the development of new regularization techniques for ML models based on the identified error-bound terms, improving their robustness and reducing error propagation.

The neural ordinary differential equation (ODE) framework has shown promise in developing accelerated surrogate models for complex systems described by partial differential equations (PDEs). In PDE-based systems, neural ODE strategies use a two-step approach for acceleration: a nonlinear dimensionality reduction via an autoencoder, and a time integration step through a neural-network based model (neural ODE). This study explores the effectiveness of autoencoder-based neural ODE strategies for advection-dominated PDEs. It includes predictive demonstrations and delves into the sources of model acceleration, focusing on how neural ODEs achieve this. The research quantifies the impact of autoencoder and neural ODE components on system time-scales through eigenvalue analysis of dynamical system Jacobians. It examines how various training parameters, like training methods, latent space dimensionality, and training trajectory length, affect model accuracy and latent time-scales. Particularly, it highlights the influence of training trajectory length on neural ODE time-scales, noting that longer trajectories enhance limiting time-scales, with optimal neural ODEs capturing the largest time-scales of the actual system. The study conducts demonstrations on two distinct unsteady fluid dynamics settings influenced by advection: the Kuramoto-Sivashinsky equations and Hydrogen-Air channel detonations, using the compressible reacting Navier--Stokes equations.

Data-driven surrogate modeling has surged in capability in recent years with the emergence of graph neural networks (GNNs), which can operate directly on mesh-based representations of data. The goal of this work is to introduce an interpretability enhancement procedure for GNNs, with application to unstructured mesh-based fluid dynamics modeling. Given a black-box baseline GNN model, the end result is an interpretable GNN model that isolates regions in physical space, corresponding to sub-graphs, that are intrinsically linked to the forecasting task while retaining the predictive capability of the baseline. These structures identified by the interpretable GNNs are adaptively produced in the forward pass and serve as explainable links between the baseline model architecture, the optimization goal, and known problem-specific physics. Additionally, through a regularization procedure, the interpretable GNNs can also be used to identify, during inference, graph nodes that correspond to a majority of the anticipated forecasting error, adding a novel interpretable error-tagging capability to baseline models. Demonstrations are performed using unstructured flow field data sourced from flow over a backward-facing step at high Reynolds numbers, with geometry extrapolations demonstrated for ramp and wall-mounted cube configurations.

Graph neural networks (GNNs) have shown promise in learning unstructured mesh-based simulations of physical systems, including fluid dynamics. In tandem, geometric deep learning principles have informed the development of equivariant architectures respecting underlying physical symmetries. However, the effect of rotational equivariance in modeling fluids remains unclear. We build a multi-scale equivariant GNN to forecast fluid flow and study the effect of modeling invariant and non-invariant representations of the flow state. We evaluate the model performance of several equivariant and non-equivariant architectures on predicting the evolution of two fluid flows, flow around a cylinder and buoyancy-driven shear flow, to understand the effect of equivariance and invariance on data-driven modeling approaches. Our results show that modeling invariant quantities produces more accurate long-term predictions and that these invariant quantities may be learned from the velocity field using a data-driven encoder.

This work introduces Jacobian-scaled K-means (JSK-means) clustering, which is a physics-informed clustering strategy centered on the K-means framework. The method allows for the injection of underlying physical knowledge into the clustering procedure through a distance function modification: instead of leveraging conventional Euclidean distance vectors, the JSK-means procedure operates on distance vectors scaled by matrices obtained from dynamical system Jacobians evaluated at the cluster centroids. The goal of this work is to show how the JSK-means algorithm -- without modifying the input dataset -- produces clusters that capture regions of dynamical similarity, in that the clusters are redistributed towards high-sensitivity regions in phase space and are described by similarity in the source terms of samples instead of the samples themselves. The algorithm is demonstrated on a complex reacting flow simulation dataset (a channel detonation configuration), where the dynamics in the thermochemical composition space are known through the highly nonlinear and stiff Arrhenius-based chemical source terms. Interpretations of cluster partitions in both physical space and composition space reveal how JSK-means shifts clusters produced by standard K-means towards regions of high chemical sensitivity (e.g., towards regions of peak heat release rate near the detonation reaction zone). The findings presented here illustrate the benefits of utilizing Jacobian-scaled distances in clustering techniques, and the JSK-means method in particular displays promising potential for improving former partition-based modeling strategies in reacting flow (and other multi-physics) applications.

The goal of this work is to address two limitations in autoencoder-based models: latent space interpretability and compatibility with unstructured meshes. This is accomplished here with the development of a novel graph neural network (GNN) autoencoding architecture with demonstrations on complex fluid flow applications. To address the first goal of interpretability, the GNN autoencoder achieves reduction in the number nodes in the encoding stage through an adaptive graph reduction procedure. This reduction procedure essentially amounts to flowfield-conditioned node sampling and sensor identification, and produces interpretable latent graph representations tailored to the flowfield reconstruction task in the form of so-called masked fields. These masked fields allow the user to (a) visualize where in physical space a given latent graph is active, and (b) interpret the time-evolution of the latent graph connectivity in accordance with the time-evolution of unsteady flow features (e.g. recirculation zones, shear layers) in the domain. To address the goal of unstructured mesh compatibility, the autoencoding architecture utilizes a series of multi-scale message passing (MMP) layers, each of which models information exchange among node neighborhoods at various lengthscales. The MMP layer, which augments standard single-scale message passing with learnable coarsening operations, allows the decoder to more efficiently reconstruct the flowfield from the identified regions in the masked fields. Analysis of latent graphs produced by the autoencoder for various model settings are conducted using using unstructured snapshot data sourced from large-eddy simulations in a backward-facing step (BFS) flow configuration with an OpenFOAM-based flow solver at high Reynolds numbers., Comment: 30 pages, 17 figures. Correction: Fixed authorship

For safety purposes, reliable reignition of aircraft engines in the event of flame blow-out is a critical requirement. Typically, an external ignition source in the form of a spark is used to achieve a stable flame in the combustor. However, such forced turbulent ignition may not always successfully relight the combustor, mainly because the state of the combustor cannot be precisely determined. Uncertainty in the turbulent flow inside the combustor, inflow conditions, and spark discharge characteristics can lead to variability in sparking outcomes even for nominally identical operating conditions. Prior studies have shown that of all the uncertain parameters, turbulence is often dominant and can drastically alter ignition behavior. For instance, even when different fuels have similar ignition delay times, their ignition behavior in practical systems can be completely different. In practical operating conditions, it is challenging to understand why ignition fails and how much variation in outcomes can be expected. The focus of this work is to understand relight variability induced by turbulence for two different aircraft fuels, namely Jet-A and a variant named C1. A detailed, previously developed simulation approach is used to generate a large number of successful and failed ignition events. Using this data, the cause of misfire is evaluated based on a discriminant analysis that delineates the difference between turbulent initial conditions that lead to ignition or failure. It was determined that the cause of ignition failure is different for the two fuels. While it was found that Jet-A is influenced by fuel entrainment, C1 was found to be more sensitive to small scale turbulence features. A larger variability is found in the ignition modes of C1, which can be subject to extreme events induced by kernel breakdown.

A method which casts the chemical source term computation into an artificial neural network (ANN)-inspired form is presented. This approach is well-suited for use on emerging supercomputing platforms that rely on graphical processing units (GPUs). The resulting equations allow for a GPU-friendly matrix-multiplication based source term estimation where the leading dimension (batch size) can be interpreted as the number of chemically reacting cells in the domain; as such, the approach can be readily adapted in high-fidelity solvers for which an MPI rank offloads the source term computation task for a given number of cells to the GPU. Though the exact ANN-inspired recasting shown here is optimal for GPU environments as-is, this interpretation allows the user to replace portions of the exact routine with trained, so-called approximate ANNs, where the goal of these approximate ANNs is to increase computational efficiency over the exact routine counterparts. Note that the main objective of this paper is not to use machine learning for developing models, but rather to represent chemical kinetics using the ANN framework. The end result is that little-to-no training is needed, and the GPU-friendly structure of the ANN formulation during the source term computation is preserved. The method is demonstrated using chemical mechanisms of varying complexity on both 0-D auto-ignition and 1-D channel detonation problems, and the details of performance on GPUs are explored.

Simultaneous measurements, such as the combination of particle image velocimetry (PIV) for velocity fields with planar laser induced fluorescence (PLIF) for species fields, are widely used in experimental turbulent combustion applications for the analysis of a plethora of complex physical processes. Such physical analyses are driven by the interpretation of spatial correlations between these fields by the experimenter. However, these correlations also imply some amount of intrinsic redundancy; the simultaneous fields contain overlapping information content. The goal of this work lies in the quantitative extraction of this overlapping information content in simultaneous field measurements. Specifically, the amount of PIV information contained in simultaneously measured OH-PLIF fields in the domain of a swirl-stabilized combustor is sought. This task is accomplished using machine learning techniques based on artificial neural networks designed to optimize PLIF-to-PIV mappings. It was found that most of the velocity information content could be retrieved when considering linear combinations of neighborhoods of OH-PLIF signal spanning roughly two integral lengthscales (half of the considered domain), and that PLIF signal interactions contained in smaller, local regions (less than half of the domain) contained no PIV information. Further, by visualizing the coherent structures contained within the neural network parameters, the role of multi-scale interactions related to velocity field retrieval from the OH-PLIF signal became more apparent. Overall, this study reveals a useful pathway (in the form of overlapping information content extraction) to develop diagnostic tools that capture more information using the same experimental resources by minimizing redundancy.

This work utilizes data-driven methods to morph a series of time-resolved experimental OH-PLIF images into corresponding three-component planar PIV fields in the closed domain of a premixed swirl combustor. The task is carried out with a fully convolutional network, which is a type of convolutional neural network (CNN) used in many applications in machine learning, alongside an existing experimental dataset which consists of simultaneous OH-PLIF and PIV measurements in both attached and detached flame regimes. Two types of models are compared: 1) a global CNN which is trained using images from the entire domain, and 2) a set of local CNNs, which are trained only on individual sections of the domain. The locally trained models show improvement in creating mappings in the detached regime over the global models. A comparison between model performance in attached and detached regimes shows that the CNNs are much more accurate across the board in creating velocity fields for attached flames. Inclusion of time history in the PLIF input resulted in small noticeable improvement on average, which could imply a greater physical role of instantaneous spatial correlations in the decoding process over temporal dependencies from the perspective of the CNN. Additionally, the performance of local models trained to produce mappings in one section of the domain is tested on other, unexplored sections of the domain. Interestingly, local CNN performance on unseen domain regions revealed the models' ability to utilize symmetry and antisymmetry in the velocity field. Ultimately, this work shows the powerful ability of the CNN to decode the three-dimensional PIV fields from input OH-PLIF images, providing a potential groundwork for a very useful tool for experimental configurations in which accessibility of forms of simultaneous measurements are limited.

In lean premixed combustors, flame stabilization is an important operational concern that can affect efficiency, robustness and pollutant formation. The focus of this paper is on flame lift-off and re-attachment to the nozzle of a swirl combustor. Using time-resolved experimental measurements, a data-driven approach known as cluster-based reduced order modeling (CROM) is employed to 1) isolate key flow patterns and their sequence during the flame transitions, and 2) formulate a forecasting model to predict the flame instability. The flow patterns isolated by the CROM methodology confirm some of the experimental conclusions about the flame transition mechanism. In particular, CROM highlights the key role of the precessing vortex core (PVC) in the flame detachment process in an unsupervised manner. For the attachment process, strong flow recirculation far from the nozzle appears to drive the flame upstream, thus initiating re-attachment. Different data-types (velocity field, OH concentration) were processed by the modeling tool, and the predictive capabilities of these different models are also compared. It was found that the swirling velocity possesses the best predictive properties, which gives a supplemental argument for the role of the PVC in causing the flame transition. The model is tested against unseen data and successfully predicts the probability of flame transition (both detachment and attachment) when trained with swirling velocity with minimal user input. The model trained with OH-PLIF data was only successful at predicting the flame attachment, which implies that different physical mechanisms are present for different types of flame transition. Overall, these aspects show the great potential of data-driven methods, particularly probabilistic forecasting techniques, in analyzing and predicting large-scale features in complex turbulent combustion problems., Comment: Preprint accepted to Combustion Theory and Modelling