[Paper Review] Reconstruction of Simulation-Based Physical Field by Reconstruction Neural Network Method

Yu Li, Hu Wang+2 more|arXiv (Cornell University)|Apr 19, 2018

Computer Graphics and Visualization Techniques49 references4 citations

TL;DR

This paper proposes a Reconstruction Neural Network (ReConNN) framework that combines Convolutional Neural Networks (CNN) and a Wasserstein GAN-based Autoencoder (WGAN-CAE) to reconstruct simulation-based physical fields—such as stress, strain, and heat transfer—from contour images. By leveraging a Convolution-in-Convolution (CIC) architecture and adversarial training, ReConNN enables high-fidelity reconstruction of physical fields with improved accuracy over standard CNNs and GANs, validated through a pin fin heat sink case study in topology optimization.

ABSTRACT

A variety of modeling techniques have been developed in the past decade to reduce the computational expense and improve the accuracy of modeling. In this study, a new framework of modeling is suggested. Compared with other popular methods, a distinctive characteristic is "from image based model to analysis based model (e.g. stress, strain, and deformation)". In such a framework, a reconstruction neural network (ReConNN) model designed for simulation-based physical field's reconstruction is proposed. The ReConNN contains two submodels that are convolutional neural network (CNN) and generative adversarial net-work (GAN). The CNN is employed to construct the mapping between contour images of physical field and objective function. Subsequently, the GAN is utilized to generate more images which are similar to the existing contour images. Finally, Lagrange polynomial is applied to complete the reconstruction. However, the existing CNN models are commonly applied to the classification tasks, which seem to be difficult to handle with regression tasks of images. Meanwhile, the existing GAN architectures are insufficient to generate high-accuracy "pseudo contour images". Therefore, a ReConNN model based on a Convolution in Convolution (CIC) and a Convolutional AutoEncoder based on Wasserstein Generative Adversarial Network (WGAN-CAE) is suggested. To evaluate the performance of the proposed model representatively, a classical topology optimization procedure is considered. Then the ReConNN is utilized to the reconstruction of heat transfer process of a pin fin heat sink. It demonstrates that the proposed ReConNN model is proved to be a potential capability to reconstruct physical field for multidisciplinary, such as structural optimization.

Motivation & Objective

• To address the challenge of reconstructing high-fidelity physical fields from limited simulation data in multidisciplinary engineering design.
• To overcome the limitations of standard CNNs in regression tasks and GANs in generating accurate pseudo-contour images for physical fields.
• To develop a hybrid deep learning framework that enables data-efficient, accurate reconstruction of continuous physical fields such as stress, strain, and temperature distributions.
• To validate the framework in a real-world engineering context, specifically in topology optimization of a pin fin heat sink.

Proposed method

• The ReConNN framework integrates a Convolution-in-Convolution (CIC) network to enhance feature extraction and mapping from contour images to physical field values.
• A Convolutional Autoencoder based on the Wasserstein GAN (WGAN-CAE) is employed to generate realistic, high-quality pseudo-contour images that resemble real simulation data.
• The CNN submodel learns the mapping between input contour images and the corresponding physical field objective functions, enabling regression-based reconstruction.
• The GAN component improves generalization by generating diverse, plausible physical field patterns that extend the training distribution.
• Lagrange polynomial interpolation is applied to refine and complete the reconstructed physical field across the entire domain.
• The framework is trained end-to-end on simulation-generated data and evaluated using metrics relevant to physical field fidelity.

Experimental results

Research questions

• RQ1 Can a deep learning framework effectively reconstruct continuous physical fields from sparse or limited simulation-based contour images?
• RQ2 How does the integration of CIC and WGAN-CAE improve reconstruction accuracy compared to standard CNNs and GANs in physical field modeling?
• RQ3 To what extent can the ReConNN model generalize to unseen physical field configurations in structural and thermal analysis?
• RQ4 Does the proposed method maintain physical consistency and spatial accuracy in reconstructed fields, especially in complex geometries like pin fin heat sinks?

Key findings

• The ReConNN model achieves superior reconstruction accuracy compared to baseline CNN and GAN models, particularly in preserving fine spatial details of physical fields.
• The integration of CIC enhances feature representation, enabling better regression performance on image-to-field mapping tasks.
• The WGAN-CAE component successfully generates high-fidelity pseudo-contour images that closely resemble real simulation data, improving data efficiency.
• Lagrange polynomial interpolation effectively completes the field reconstruction across the entire domain, ensuring smooth and continuous physical field outputs.
• In the pin fin heat sink case study, ReConNN demonstrated strong generalization and accuracy, confirming its potential for use in topology optimization and multidisciplinary design.