Documents published in Scipedia

• Using conservation laws to infer deep learning model accuracy of Richtmyer-Meshkov instabilities

  C. Jekel, D. Sterbentz, S. Aubry, Y. Choi, D. White, J. Belof

  eccomas2022.

  
  

  Abstract

  Richtmyer-Meshkov Instability (RMI) is a complicated phenomenon that occurs when a shockwave passes through a perturbed interface. Over a thousand hydrodynamic simulations were performed to study the formation of RMI for a parameterized high velocity impact. Deep learning was used to learn the temporal mapping of initial geometric perturbations to the full-field hydrodynamic solutions of density and velocity. The continuity equation was used to include physical information into the loss function, however only resulted in very minor improvements at the cost of additional training complexity. Predictions from the deep learning model appear to accurately capture temporal RMI formations for a variety of geometric conditions within the domain. First principle physical laws were investigated to infer the accuracy of the model's predictive capability. While the continuity equation appeared to show no correlation with the accuracy of the model, conservation of mass and momentum were weakly correlated with accuracy. Since conservation laws can be quickly calculated from the deep learning model, they may be useful in applications where a relative accuracy measure is needed.

  Abstract
  Richtmyer-Meshkov Instability (RMI) is a complicated phenomenon that occurs when a shockwave passes through a perturbed interface. Over a thousand hydrodynamic simulations [...]

  • 8
  •  read

• Vectorial limitation for multislope MUSCL schemes

  A. Tételin, C. Le Touze, P. Villedieu

  eccomas2022.

  
  

  Abstract

  In finite volume schemes with MUSCL interpolation of scalar variables at the faces of control volumes, a slope limiting function is used in order to prevent non-physical oscillations of the solution. More particularly, these functions are designed to ensure a certain monotonicity criterion at each face of the control volume, criterion which then ensures a stability property of the scheme. For vectorial variables, these slope limiting functions are generally applied componentwise, but this may result in a frame-dependance, as well as a loss of accuracy due to false detection of extrema. In this paper, a new vectorial interpolation method is introduced, which is frame-invariant, second-order accurate and stable in a sense that will be defined.

  Abstract
  In finite volume schemes with MUSCL interpolation of scalar variables at the faces of control volumes, a slope limiting function is used in order to prevent non-physical oscillations [...]

  • 3
  •  read

• Remaining Useful Life prediction with a Deep Self-Supervised Learning Approach

  A. Akrim, C. Gogu, R. Vingerhoeds, M. Salaün

  eccomas2022.

  
  

  Abstract

  With the increasing availability of data for Prognostics and Health Management (PHM), Deep Learning (DL) techniques are now the subject of considerable attention in Prognostics for Predictive Maintenance, achieving more accurate Remaining Useful Life (RUL) predictions. However, one of the major challenges for DL techniques resides in the difficulty of obtaining large amounts of labeled data on industrial systems. To overcome this lack of labeled data, an emerging learning technique is considered in this work : Self-Supervised Learning, a sub-category of unsupervised learning approaches. This paper aims to investigate whether pre-training DL models in a self-supervised way on unlabeled sensors data can be useful for downstream tasks in PHM (i.e. RUL estimation) with only limited amount of labelled data. A synthetic dataset composed of strain data is used. Results show that the self-supervised pretrained models significantly outperform the non pre-trained models in downstream Remaining Useful Life (RUL) prediction task, showing promising results in prognostic tasks when only limited labeled data is available.

  Abstract
  With the increasing availability of data for Prognostics and Health Management (PHM), Deep Learning (DL) techniques are now the subject of considerable attention in Prognostics [...]

  • 8
  •  read

• Pros and cons of various equivalent frame models for nonlinear analysis of URM buildings

  A. Shabani, M. Kioumarsi

  eccomas2022.

  
  

  Abstract

  Brick masonry is considered as one of the old construction materials, and several cultural heritage assets are made of unreinforced masonry (URM), which is susceptible to earthquakes due to its brittle behavior. The equivalent frame method (EFM) is a nonlinear modeling method widely utilized for the seismic analysis of URM buildings with lower computational efforts than finite and discrete element methods. In this study, three macroelements, including the unified method (UM), composite spring method (CSM), and double modified multiple vertical line element model (DM-MVLEM), were utilized to model three case studies. The first case study is a full-scale two-story URM wall that was tested by applying the cyclic prescribed displacements, and two other case studies were developed by changing the configuration of openings. The second case study is with short piers, and weak spandrels exist in the third model. The efficiency of the methods in terms of the accuracy of the pushover results, prediction of damage patterns, and duration of the incremental dynamic analysis (IDA) are discussed. Finally, seismic fragility curves are provided to compare the IDA results.

  Abstract
  Brick masonry is considered as one of the old construction materials, and several cultural heritage assets are made of unreinforced masonry (URM), which is susceptible to [...]

  • 8
  •  read

• Simulation of reacting flows using artificial neural networks: application to multi-regime combustion

  C. Mehl, D. Aubagnac-Karkar

  eccomas2022.

  
  

  Abstract

  The simulation of reactive flows is a major challenge in several industrial sectors, such as aeronautics or energy production. The coupling between fluid dynamics and chemistry comes however at a cost, as chemical processes involve a wide range of spatial and temporal scales. The resulting equations are stiff and require specific, and expensive, numerical methods. The use of machine learning to estimate the reaction rates has been recently proposed. In particular, Artificial Neural Networks (ANN) have the ability to perform interpolation on high-dimensional data and are thus particularly adapted to chemistry problems. A major issue is then to select an appropriate database on which to train the ANN. It must: (i) be representative of the targeted application; (ii) be sufficiently quick to generate. A promising strategy is to use 0-D stochastics reactors, which mimic reactive and mixing processes in systems while being cheap to compute. This methodology has been successfully applied to non-premixed combustion in the literature. In the present work, the aim is to investigate the ability of the 0D stochastic reactors to be used as a database for a wider range of combustion systems. More specifically, the focus will be on the ability to predict auto-ignition followed by premixed flame propagation. To that purpose, a 2-D turbulent case involving the auto-ignition of a hotspot in a hydrogen/air mixture and the subsequent propagation of a premixed flame is proposed. An ANN model based on stochastic reactors is then built and tested on (i) a 0-D auto-ignition case; (ii) a 1-D laminar premixed flame propagation; (ii) the full 2-D turbulent configuration. Using adequate data transformation at the input and output of the neural network, accurate results are obtained, highlighting the ability of the proposed strategy to deal with a large range of combustion applications.

  Abstract
  The simulation of reactive flows is a major challenge in several industrial sectors, such as aeronautics or energy production. The coupling between fluid dynamics and chemistry [...]

  • 3
  •  read

• Scalability analysis and performance modelling of layer-parallel training of deep residual networks using a non-linear multigrid-in-time algorithm

  C. Datar, H. Köstler

  eccomas2022.

  
  

  Abstract

  Replacing the traditional forward and backward passes in a residual network with a Multigrid-Reduction-in-Time (MGRIT) algorithm paves the way for exploiting parallelism across the layer dimension. In this paper, we evaluate the layer-parallel MGRIT algorithm with respect to convergence, scalability, and performance on regression problems. Specifically, we demonstrate that a few MGRIT iterations solve the systems of equations corresponding to the forward and backward passes in ResNets up to reasonable tolerances. We also demonstrate that the MGRIT algorithm breaks the scalability barrier created by the sequential propagation of data during the forward and backward passes. Moreover, we show that ResNet training using the layer-parallel algorithm significantly reduces the training time compared to the layer-serial algorithm on two non-linear regression tasks. We observe much more efficient training loss curves using layer-parallel ResNets as compared to the layer-serial ResNets on two regression tasks. We hypothesize that the error stemming from approximately solving the forward and backward pass systems using the MGRIT algorithm helps the optimization algorithm escape flat saddle-point-like plateaus or local minima on the optimization landscape. We validate this by illustrating that artificially injecting noise in a typical forward or backward propagation, allows the optimizer to escape a saddle-point-like plateau at network initialization.

  Abstract
  Replacing the traditional forward and backward passes in a residual network with a Multigrid-Reduction-in-Time (MGRIT) algorithm paves the way for exploiting parallelism across [...]

  • 3
  •  read

• Lessons Learnt from Chimera Method Application to a Deploying Krueger Device

  A. Hasabnis, H. Maseland, F. Moens

  eccomas2022.

  
  

  Abstract

  The Chimera method [1] is an established method for simulation of overlapping grids. Meshing parts independently has made this method popular for complex geometries as well as moving bodies like propellers and rotors (e.g. [2]) or control surfaces (e.g. [3]). It is thus a promising method to simulate deflecting high-lift systems. The motion of the Krueger flap – as the most promising leading edge high-lift system device for laminar wing technology – is characterized by a relatively large movement (about 140 deg deflection) at a relatively high deflection speed (up to 200 deg/s) compared to classical leading edge devices. In terms of simulation, the grid properties of the overlapping mesh regions vary throughout the motion from a nearly sealed retracted position to a gapped flow in fully deflected position comparable to a slat device. This expects dynamic effects may get dominant and a valid simulation of this flow is needed for proper design and analysis. In the frame of the UHURA project1 , several partners applied their CFD capabilities based on Chimera in order to validate the method for this specific application in comparison to wind tunnel tests. The presentation outlines the different Chimera approaches ranging from structured/2D to hybrid/3D in steady and unsteady simulations for the different type of setups investigated, namely straight and swept wing with full-span and part-span Krueger flap. It summarizes common challenges and best practice for application of the Chimera approach for such a device. 

  Abstract
  The Chimera method [1] is an established method for simulation of overlapping grids. Meshing parts independently has made this method popular for complex geometries as well [...]

  • 15
  •  read

• Computational modelling of plasma electrolytic oxidation process induced damage in extruded Mg material

  E. Gazenbiller, S. Mansoor, N. Konchakova, M. Serdechnova, C. Blawert, M. Zheludkevich, D. Höche

  eccomas2022.

  
  

  Abstract

  Magnesium (Mg) alloys are an attractive constructive material due to their light weight and high mechanical strength. Plasma electrolyte oxidation (PEO) treatment of Mg alloys creates a thin ceramic coating with protective effects against mechanical wear and corrosion. The coating properties like its porosity and thickness can be adjusted by PEO process parameters and at the same time affects the material behaviour under tensile strength. In this work, dedicated slow-strain rate experiments of differently PEO coated Mg alloy dog-bone shaped specimen were conducted and the coating porosity, thickness and crack spacing were analyzed in order to deduce a predictive Finite Element Method (FEM) damage model. The results indicate that the thicker, more porous coatings lead to material failure at smaller strains in plastic regions. The effect can be implemented via partial differential equation into the FEM model.

  Abstract
  Magnesium (Mg) alloys are an attractive constructive material due to their light weight and high mechanical strength. Plasma electrolyte oxidation (PEO) treatment of Mg alloys [...]

  • 6
  •  read

• Parallel Finite Volume Code for Plasma with Unstructured Adaptive Mesh Refinement

  I. Kissami, S. maazioui, F. Benkhaldoun

  eccomas2022.

  
  

  Abstract

  The present paper describes a parallel unstructured-mesh Plasma simulation code based on Finite Volume method. The code dynamically refines and coarses mesh for accurate resolution of the different features regarding the electron density. Our purpose is to examine the performance of a new Parallel Adaptive Mesh Refinement (PAMR) procedure introduced on the ADAPT platform, which resolves of a relatively complicated system coupling the flow partial differential equations to the Poisson's equation. The implementation deals with the MUMPS parallel multi-frontal direct solver and mesh partitioning methods using METIS to improve the performance of the framework. The standard MPI is used to establish communication between processors. Performance analysis of the PAMR procedure shows the efficiency and the potential of the method for the propagation equations of ionization waves.

  Abstract
  The present paper describes a parallel unstructured-mesh Plasma simulation code based on Finite Volume method. The code dynamically refines and coarses mesh for accurate resolution [...]

  • 21
  •  read

• PREDICTION OF THE BIOMASS PARTICLES THROUGH THE PHYSICS-INFORMED NEURAL NETWORK

  F. darlik, P. Adhav, B. Peters

  eccomas2022.

  
  

  Abstract

  Woody biomass energy is a kind of renewable energy that contributes to the reduction of greenhouse gas emissions, the creation of healthier forests, and the reduction of wildfire danger. Generally speaking, simulations of the motion of biomass particles are a time-consuming process due to a large number of particles and required simulation time. We used a physicsinformed neural network (PINN) model to predict the motion of particles by including their equations of motion to reconstruct the velocity fields and reduce the processing effort. compare to the discrete element methods, the PINNs methods have the advantage of predicting the velocity fields without the knowledge of the simulation's boundary and initial conditions as well as geometry. It has shown that the proposed model has reliable prediction results with a mean percentage error in time less than 1 percent.

  Abstract
  Woody biomass energy is a kind of renewable energy that contributes to the reduction of greenhouse gas emissions, the creation of healthier forests, and the reduction of wildfire [...]

  • 4
  •  read