Documents published in Scipedia

• Data-driven modeling of complex mechanical components for integration in system-level simulations

  S. Vanpaemel, N. Kutz, S. Brunton

  WCCM2024.

  
  

  Abstract

  This contribution presents a data-driven approach featuring a physics-inspired neural network structure for modeling complex components in mecha(tro)nic systems. In the present approach, gated recurrent units (GRUs) are employed to approximate the ordinary differential equations (ODEs) describing the system’s states over time, followed by a deep feedforward neural network (FFNN) mapping these states to a target variable. The networks are shown to predict a latent space capable of modeling the underlying dynamics, without the need for measuring the full state vector and only relying on input-output measurements. Subsequently it is shown that a nonlinear coordinate transformation exists between the latent space of the network and the states obtained from the reference ODE integration (simulation). To have a verification of the network’s performance, it is applied to simulation-based data of an academic example for which the states and equations are known beforehand. Furthermore, the methodology is also applied to real measurement data from an INSTRON testing system capturing shock damper and bushing dynamic behaviour. Lastly, it is demonstrated that an ODE expression can be extracted from the trained network. This feature allows seamless integration of these networks into variable time-step, system-level simulation software.

  Abstract
  This contribution presents a data-driven approach featuring a physics-inspired neural network structure for modeling complex components in mecha(tro)nic systems. In the present [...]

  • 49
  •  read

• Data-efficient one-step mechanical design of composites using generative AI

  M. Masrouri, Z. Qin

  WCCM2024.

  
  

  Abstract

  The distribution of material phases is crucial to determine the composite's mechanical properties. While the entire structure-mechanics relationship of highly ordered material distributions can be studied with a finite number of cases, this relationship is challenging to reveal for complex irregular distributions, preventing the design of such material structures from meeting specific mechanical requirements. The noticeable developments of artificial intelligence algorithms in material design enable the discovery of hidden structure-mechanics correlations, which is essential for designing composites of complex structures. It is intriguing how these tools can assist composite design. Here, we focus on the rapid generation of complex irregular composite structures and the stress distribution in loading. We find that generative AI, enabled through fine-tuned Low-Rank Adaptation models, can be trained with a few inputs to generate synthetic composite structures and the corresponding von Mises stress distribution. The results show that this technique is convenient in generating massive composite designs with useful mechanical information that dictates stiffness, fracture, and robustness of the material with one model, and such must be done by several different experimental or simulation tests. This research offers valuable insights for improving composite design to expand the design space and automatic screening of composite designs for improved mechanical functions.

  Abstract
  The distribution of material phases is crucial to determine the composite's mechanical properties. While the entire structure-mechanics relationship of highly ordered material [...]

  • 101
  •  read

• A dynamic weighted loss function for enhancing the performance of neural networks

  C. Mang, A. Tahmasebi Moradi, D. Danan, M. Yagoubi

  WCCM2024.

  
  

  Abstract

  In machine learning process, hyper parameters are chosen in a way to decrease the prediction error and improve the convergence. However, the optimized hyper parameters have a limit in terms of enhancing the performance of the neural networks. In this work, the datasets used for the numerical experiments arise from the resolution of partial differential equations (PDE) defined on a spatial domain. We propose a DYNAmic WEIghted Loss (DYNAWEIL) function-based approach for neural networks that are used to learn these PDE’s solutions. This a two-step process: first we train for a few numbers of epochs in a classical way then the dynamic weighted loss function replaces the classical loss function by leveraging the information from past training error histories. To validate this method, we carry out numerical experiments with different neural networks on datasets arising on two different physics: Goldstein equation [1] and radiative transfer equation [2]. Thus, in order to demonstrate the relevance of this approach, we provide a comparison among a neural network model using a classical loss function, with and without hyper parameters optimization, and a dynamic weighted loss function for both versions.

  Abstract
  In machine learning process, hyper parameters are chosen in a way to decrease the prediction error and improve the convergence. However, the optimized hyper parameters have [...]

  • 85
  •  read

• Prediction of unsteady heat transfer of temperature field on a circuit board using sub-voxels input data structure

  T. Tsukiji, Y. Wada, Y. Iwata, M. Irikiin

  WCCM2024.

  
  

  Abstract

  In this study, we propose a sub-voxel learning method based on a Neural Operator and predict the thermal temperature field on a circuit board in unsteady heat conduction. CAE analysis reduces the cost of experiments using prototypes in the design phase. However, the computational cost is high because the number of trials increases due to changes in analysis conditions. Predictions made by machine learning are less accurate than those made by CAE analysis, but they can significantly reduce computational costs. In general, machine learning is difficult to extrapolate. Thus, the concept of a Neural Operator has attracted attention as one machine learning method incorporating physical laws. In this study, we propose a sub-voxel learning method based on the Neural Operator, inspired by the forward Euler method, and predict the thermal temperature field on a circuit board during unsteady heat conduction. The input data is the analysis data of the current cycle step, and the output data is the data of the next cycle step. Dummy temperatures were set according to the heat generation of each IC in the input data of the 0th cycle step, and pseudo-difference values were generated to extract features. The prediction accuracy of the next cycle step was compared with and without dummy temperatures

  Abstract
  In this study, we propose a sub-voxel learning method based on a Neural Operator and predict the thermal temperature field on a circuit board in unsteady heat conduction. [...]

  • 13
  •  read

• Graph neural networks for accelerating the discrete element simulation of granular flow

  P. Zhi, Y. Wu

  WCCM2024.

  
  

  Abstract

  Granular flow is a phenomenon widely presented in both the natural and engineering fields. Here granular materials could be either solid particles, e.g. rocks, soil, and grains, or liquid particles, e.g. mud and fresh concrete mortar. Soil landslides, particle transport, and grain accumulation have been edge-cutting hot research topics. Discrete Element Method (DEM) has been regarded as one of the most important methods to simulate granular flows and to investigate discontinuous and large deformation problems. The basic principle of DEM was to view the simulated object as consisting of discrete particles, to define specific constitutive relationships for the particles, and to study the macroscopic properties of the simulated object from a microscopic perspective based on the interactions between particles. However, DEM simulations usually consume very high computational cost for particle contact searching and detection. To accelerate the computational process of discrete element simulation, the Graph Neural Network (GNN) based deep learning model was proposed in this paper. In GNNs, graph nodes and graph edges represent the particles and their interactions. The training and testing datasets were generated using an open-source software named YADE, while the neural network model was constructed using PyTorch and Deep Graph Library (DGL). Replacing the direct calculation of particle collisions in DEM with the trained neural network model, the state of the particles at the next moment could be predicted based on the current state of the particles. It significantly increased computational speed. The proposed technique was applied in various examples, such as drum rotation and hopper stacking, and its accuracy had been verified. This study established a solid foundation and provided robust support for further research and applications of granular flow simulation based on GNN

  Abstract
  Granular flow is a phenomenon widely presented in both the natural and engineering fields. Here granular materials could be either solid particles, e.g. rocks, soil, and grains, [...]

  • 139
  •  read

• Prediction of plural crack propagation using discovered PDE

  G. MURAOKA, Y. Wada

  WCCM2024.

  
  

  Abstract

  This study presents a prediction of plural crack propagation using the discovered partial differential equations. 80% of structures fracture due to fatigue failure. Therefore, the evaluation of fatigue cracks is essential. Numerical analysis is costly, and machine-learning surrogate models have been proposed. Hence, the crack propagation path and remaining life are predicted using machine learning. A dataset is obtained from the results of a crack propagation analysis using a s-version FEM combined with an automatic mesh generation technique. The input parameters are the coordinates of the four crack tips, and the output parameters are the crack propagation vector and the number of cycles of 0.25 mm. Also, physics-informed neural networks (PINNs) have been widely studied in recent years. Thus, we took inspiration from PINNs and added a regularization term of PDE discovered by AI Feynman to the loss. As a result, the loss of a validation dataset for training constrained by PDE was reduced by about 77% compared to the unconstrained loss. The error in crack length decreased from -0.50% to 0.17%.

  Abstract
  This study presents a prediction of plural crack propagation using the discovered partial differential equations. 80% of structures fracture due to fatigue failure. Therefore, [...]

  • 121
  •  read

• Deep convolutional architectures for uncertainty quantification and forecast in inundation problems

  A. Soulaïmani, Y. Kumar, P. Bhatt, M. Moosa

  WCCM2024.

  
  

  Abstract

  This article provides a summary of our latest research, where we investigate the application of data-driven deep learning methods to simulate the dynamics of physical systems that are governed by partial differential equations (PDEs). The main challenge is the long-term temporal extrapolation for fluid dynamics problems that exhibit steep gradients and discontinuities. We make use of deep learning techniques, specifically designed for time-series predictions like LSTM, TCN, and Attention mechanism, as well as CNN. These methods are employed to model the dynamics of systems primarily influenced by advection. We propose a combination of a Convolutional Autoencoder (CAE) model for data compression and a novel CNN-based for forecasts. These models take a series of high-fidelity vector solutions and predict the solutions for the following time steps using auto-regression. To reduce complexity and computational demands during both online and offline stages, we implement deep auto-encoder networks. These techniques are used to compress the high-fidelity snapshots before feeding them into the forecasting models. Our models are evaluated on numerical benchmarks, such as the 1D Burgers’ equation and Stoker’s dam-break problem, to assess their long-term predictive accuracy, even in scenarios that extrapolate beyond the training domain. The model that demonstrates the highest accuracy is subsequently used to simulate a hypothetical dam break in a river with real 2D bathymetry. Due to space constraints, only a selection of results is showcased, with additional findings available in our work [1] and the newer ones will also be presented in the talk. Our findings indicate that the proposed CNN future-step predictor offers significantly accurate forecasts in the considered spatiotemporal problems.

  Abstract
  This article provides a summary of our latest research, where we investigate the application of data-driven deep learning methods to simulate the dynamics of physical systems [...]

  • 18
  •  read

• CNN-based surrogate model and temperature prediction method using superposition principle

  M. Irikiin, Y. Iwata

  WCCM2024.

  
  

  Abstract

  A CNN-based surrogate model is being developed to accelerate CFD calculations. In order to use this surrogate model for design development, it is necessary to improve generalizability. One solution to this problem is to use the principle of superposition. For the multiple heating elements that make up the model, their temperatures are predicted by heating them individually. We devised a method to predict the temperature of the entire model by adding up these individually predicted temperature distributions. Radiation and convection phenomena, for which the superposition principle does not hold, were also considered.

  Abstract
  A CNN-based surrogate model is being developed to accelerate CFD calculations. In order to use this surrogate model for design development, it is necessary to improve generalizability. [...]

  • 8
  •  read

• Investigation of CNN-based multigrid-bidirectional networks

  Y. Iwata, Y. Inagaki, M. Irikiin

  WCCM2024.

  
  

  Abstract

  We are developing a high-speed simulation technology for physics simulations using deep learning. This technology aims to accelerate simulation time by a factor of several hundred to a thousand, significantly enhancing product performance and quality by increasing development efficiency and optimization. Currently, we are focusing on a multi-grid convolutional neural network (CNN) based architecture designed for models incorporating a combination of coarse and dense grids, addressing the challenge of model scaling and high resolution. Our previous work, reported by the Society for Computational Engineering and Science in June 2023, demonstrated the propagation of physical information from a coarse grid to a dense grid. Building on this foundation, we have now developed a technique that facilitates the propagation of physical information among multiple grids with varying resolutions. We applied this novel method to a basic temperature distribution prediction model for circuit boards and verified its high accuracy in predicting temperature distribution.

  Abstract
  We are developing a high-speed simulation technology for physics simulations using deep learning. This technology aims to accelerate simulation time by a factor of several [...]

  • 5
  •  read

• Investigation of geotribological behaviour in the soil-pile interface considering clay based on numerical and laboratory investigations

  S. Henke, P. Wiesenthal

  WCCM2024.

  
  

  Abstract

  The contact behavior between soil and structures is an important aspect in many geotechnical applications. One example is the contact between pile and soil during pile installation which especially for open-ended pipes can lead to soil plug formation. Within the present research, contact behavior between clayey soil and pile is investigated by means of numerical and laboratory experiments focusing on the contact behavior within tubular piles. First, the contact between kaolin clay and steel is experimentally investigated with respect to overburden pressure in the so-called Geo-Tribometer developed at HSU. The results of the experimental investigations show some unexpected results leading to the assumption that the contact failure surface inside the soil specimen changes with differing overburden pressure. Additional numerical simulations are carried out for better understanding of contact stress development. Second, further laboratory investigations using soil-filled tubular piles show that adhesion-like effects significantly influence the contact behavior between soil plug and internal surface of the tube. For estimation of the adhesion values, numerical simulations by means of finite element analyses are carried out showing that as expected with increasing soil’s overburden pressure adhesion effect increases. The results are finally discussed with respect to transferability from small scale in numerical and laboratory investigations toward prototype scale.

  Abstract
  The contact behavior between soil and structures is an important aspect in many geotechnical applications. One example is the contact between pile and soil during pile installation [...]

  • 18
  •  read