Abstract
Digital twins require high-quality data to accurately mirror physical systems, yet high-fidelity (HF) data is often scarce and expensive to obtain. This paper presents GAN-MDF, a novel Generative Adversarial Network (GAN) framework designed for Multi-fidelity Data Fusion (MDF). By combining a split-generator architecture—comprising a pre-trained low-fidelity (LF) block and a high-fidelity mapping block—with a "supervised-loss trick," GAN-MDF effectively fuses abundant LF data with limited HF samples. Experimental results across ten benchmark problems demonstrate that GAN-MDF significantly outperforms traditional interpolation and machine learning methods in accuracy and robustness, particularly when HF data is extremely limited.
Introduction
Digital twins serve as virtual replicas of physical entities, such as smart factories or robotic systems, enabling real-time monitoring and predictive maintenance. The efficacy of these twins depends on the quality of the underlying data. Researchers typically encounter two data types: High-Fidelity (HF) data, which is accurate but costly, and Low-Fidelity (LF) data, which is abundant but less precise.
The challenge of Multi-fidelity Data Fusion (MDF) is to integrate these disparate sources to build a high-accuracy model. Traditional methods like Kriging interpolation often rely on strict assumptions, such as Gaussian distributions or nested data structures. Standard machine learning models, on the other hand, frequently suffer from overfitting or divergence when HF samples are scarce. GAN-MDF addresses these limitations by leveraging the generative power of adversarial networks tailored for regression tasks.
Methodology
Standard GAN vs. GAN-MDF
A standard Generative Adversarial Network typically generates data from random noise, as shown in the figure below. While effective for generating realistic images, this approach lacks the precision required for scientific data fusion.
In contrast, GAN-MDF utilizes a specialized architecture designed to extract features from LF data before mapping them to HF outputs. The framework consists of two primary components:
- The Generator: Unlike the noise-driven generator in standard GANs, the GAN-MDF generator is split into two sub-networks:
- LF Block: Pre-trained on the extensive LF dataset to capture the system's underlying physical features. Its weights are frozen after pre-training.
- HF Block: Takes the original input concatenated with the features from the LF block to learn the mapping to the HF response.
- The Discriminator: Evaluates whether the output is a generated (fake) HF sample or a real HF sample.
The Supervised-Loss Trick
To overcome the inherent instability of GANs in regression tasks, the authors introduce a supervised-loss trick. A supervised loss term ($L_S$) is added to the objective function, penalizing the direct difference between the generator's output and actual HF training samples. This term acts as a guide, ensuring the generator converges toward accurate numerical values rather than just plausible distributions.
Experimental Setup and Results
The performance of GAN-MDF was evaluated against four state-of-the-art methods (co-RBF, H-kriging, LS-MFS, and Hierarchical Regression) across ten test cases, including nine synthetic benchmarks and one real-world malaria parasite rate dataset.
Key Findings
- Superior Accuracy: GAN-MDF consistently achieved lower Normalized Root Mean Square Error (NRMSE) than competing methods across most test scenarios.
- Robustness to Scarcity: The model demonstrated remarkable stability with as few as 2 to 5 HF samples. While other methods failed to converge or showed high error rates under extreme scarcity, GAN-MDF remained accurate.
- Distribution Agnostic: Unlike Gaussian process-based methods, GAN-MDF does not require specific data distribution assumptions, making it highly flexible for diverse engineering applications.
- Unnested Flexibility: The framework performs effectively even when HF samples are not a subset of the LF data, removing a common constraint in traditional MDF techniques.
Conclusion
GAN-MDF provides a robust and flexible solution for multi-fidelity data fusion in digital twins. By combining feature extraction from LF data with a supervised adversarial training process, it overcomes the limitations of data scarcity and strict mathematical assumptions. Future work will focus on the theoretical underpinnings of the model's success and refining the architecture for even higher-dimensional engineering problems.
Summary updated