CcDPM: A Continuous Conditional Diffusion Probabilistic Model for Inverse Design

Yanxuan Zhao, Peng Zhang, Guopeng Sun, Zhigong Yang, Jianqiang Chen, Yueqing WangAAAI 2024 · DOI

Introduction

Inverse design represents a transformative paradigm in engineering, shifting from iterative trial-and-error to a goal-oriented approach where performance requirements directly dictate geometric configurations. While Generative Adversarial Networks (GANs) have previously been applied to this domain, they are often hampered by three primary hurdles: discrete performance assumptions, training instabilities such as mode collapse, and difficulties in handling multi-point design constraints—where a single design must satisfy performance targets across varying operational conditions.

Methodology: The CcDPM Framework

To address these challenges, we introduce the Continuous Conditional Diffusion Probabilistic Model (CcDPM). This framework leverages the inherent stability of diffusion models while introducing specific innovations for continuous engineering data.

Vicinal Risk Minimization (VRM) and k-d Tree Acceleration

The core of CcDPM's training lies in Vicinal Risk Minimization (VRM). Unlike standard conditional models that treat performance labels as exact points, VRM assumes a design is representative of a "vicinity" in the performance space, modeled via a Gaussian kernel. This smoothing allows the model to learn the underlying physical relationships more effectively than models that treat performance metrics as rigid, discrete categories.

To make this computationally feasible, especially on large-scale datasets, we implement a k-d tree acceleration strategy. This reduces the complexity of searching for neighboring data points from $O(N)$ to $O(\log N)$, enabling efficient training on millions of data points.

The training architecture of CcDPM, highlighting the integration of the k-d tree for efficient vicinal loss calculation.

Multi-point Design Sampling

CcDPM is uniquely capable of multi-point design. By utilizing an estimator-free guidance method (similar to classifier-free guidance in image generation), the model can simultaneously process multiple performance requirements under different working conditions (e.g., lift at different Mach numbers) to generate a single optimized design that satisfies all conditions at once.

Experimental Evaluation

The model was rigorously tested across four datasets: a synthetic uneven distribution, the Xfoil airfoil dataset (48,503 samples), a high-fidelity CFD dataset (10,366 samples), and a massive Rocket dataset containing over 3.3 million data points.

Accuracy and Diversity

CcDPM consistently achieved the lowest label error (the difference between requested and actual performance) across all datasets, outperforming state-of-the-art baselines like CcGAN and PcDGAN. Qualitatively, the model demonstrates superior robustness in generating valid geometries across the entire performance spectrum, avoiding the mode collapse frequently seen in GAN-based approaches.

Qualitative comparison of generated airfoil shapes across different performance requirements. CcDPM maintains structural integrity and diversity where GAN-based models often fail or collapse.

Training Efficiency

The k-d tree optimization proved critical for scalability. On the Rocket dataset, training time was reduced from a projected 530 hours to just 1.5 hours, representing a 353x speedup. This efficiency makes continuous diffusion models practical for industrial-scale engineering problems that were previously prohibitively slow.

Conclusion and Future Work

CcDPM establishes a new state-of-the-art for engineering inverse design by combining the stability of diffusion models with continuous label handling and efficient spatial indexing. While currently limited by geometric parameterization dependencies (such as Bézier curves), future iterations focusing on latent space diffusion promise to extend these benefits to complex topology optimization and non-parameterized geometry generation.

Summary updated

BibTeX
@inproceedings{zhao2024ccdpm,
  author = {Yanxuan Zhao and Peng Zhang and Guopeng Sun and Zhigong Yang and Jianqiang Chen and Yueqing Wang},
  title = {{CcDPM: A Continuous Conditional Diffusion Probabilistic Model for Inverse Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i15.29647},
}
Figure