Abstract
High-fidelity physical simulations such as Finite Element Analysis (FEA) are computationally intensive, often limiting the availability of large datasets for deep learning applications. This paper presents ReConNN (Reconstruction Neural Network), a framework that integrates a Convolution in Convolution (CIC) regression model with a WGAN-CAE generative architecture. ReConNN reconstructs high-fidelity physical fields from limited simulation data by generating accurate pseudo-contour images. Experimental results on topology optimization and heat transfer problems demonstrate that ReConNN significantly outperforms standard CNN architectures in regression accuracy and generative stability, enabling the expansion of sparse simulation frames into dense, high-resolution sequences.
Introduction
In modern engineering, high-fidelity simulations like Finite Element Analysis (FEA) are essential for physical analysis but come with a high computational cost. Traditional metamodels, such as Kriging or Support Vector Regression, often simplify these complex simulations by mapping input parameters to single scalar outputs, thereby discarding critical spatial information. While Convolutional Neural Networks (CNNs) are adept at processing spatial data, they are typically optimized for classification rather than the high-precision regression required for physical fields. Standard architectures often plateau when tasked with capturing the nuanced gradients of structural stress or heat transfer.
ReConNN addresses these challenges by shifting the modeling paradigm from "analysis-to-objective" to "image-to-analysis." By treating simulation results as images, the framework leverages computer vision advancements to reconstruct complex physical fields more accurately and with fewer high-fidelity simulation runs.
Methods
The ReConNN framework employs a two-pronged architectural approach:
Convolution in Convolution (CIC)
To improve regression accuracy, the CIC submodel utilizes an "image incision" technique. The original high-resolution image is divided into 30 sub-images, each processed by its own specialized sub-CNN. This architecture significantly increases the number of trainable parameters without a proportional increase in computational cost, allowing the model to focus on local features that drive physical behavior.
WGAN-CAE Generative Model
To augment limited simulation data, the authors developed a hybrid generative model combining a Wasserstein GAN (WGAN) with a Convolutional AutoEncoder (CAE). The CAE compresses complex contour images into a stable 8x8 latent space. The WGAN then learns to generate new samples within this reduced dimensionality, which are subsequently "inflated" back to the original resolution by the CAE's decoder. This approach avoids the instability common when training GANs on high-resolution raw pixels.
Datasets and Evaluation
The framework was evaluated on two engineering benchmarks:
- Topology Optimization: A dataset of 22,000 contour images representing structural bridge design evolution, labeled with structural compliance.
- Heat Transfer: A 500-sample dataset documenting the cooling process of a "pin fin" heat sink.
Results and Discussion
ReConNN demonstrated superior performance in both regression and generation tasks. The CIC model achieved a Root Mean Square Error (RMSE) of 0.0354, outperforming LeNet (0.5755) and VGG (0.2346). Analysis of model weights revealed that while standard models like AlexNet often rely on bias terms (effectively "guessing"), the CIC model successfully learned underlying physical features.
In generative tasks, the WGAN-CAE achieved an Inception Score of 5.97, indicating higher diversity and physical representativeness compared to DCGAN (4.62) and standard WGAN (1.02).
Furthermore, the model successfully performed "field expansion." In the topology optimization test, it reconstructed a smooth sequence of 6,400 frames from just 64 original simulation frames, effectively filling gaps in the simulation data with high fidelity.
Conclusion
ReConNN provides a robust methodology for reconstructing physical fields from limited data by treating simulation outputs as spatial images. While the current implementation requires a substantial initial dataset to train the CAE and CIC models, it offers a path toward more efficient engineering analysis. Future research will explore extending the framework to 3D physical fields and integrating it with Digital Image Correlation (DIC) for real-world experimental reconstruction.
Summary updated