关键词: 关键词：多约束设计, 生成式设计, 扩散Transformer, 傅里叶特征嵌入器, 条件随机丢弃

Abstract: Abstract: To address the requirements of shape design based on sparse performance indicators in the early stages of aircraft conceptual design, a multi-constraint generative design method based on the Diffusion Trans-former (DiT) is proposed. Taking aerodynamic and geometric parameters such as lift-to-drag ratio, missile diameter, and wing root chord length as inputs, this method constructs a diffusion denoising network to generate continuous parameterized shapes. Regarding the condition introduction mechanism, a multi-frequency Fourier feature em-bedder is designed to process continuous scalars, and the extracted conditional features are concatenated with the initial noise sequence along the sequence length dimension to directly participate in the self-attention computation. For multi-constraint inputs, a multi-condition joint random dropout strategy is developed to apply global and inde-pendent masking operations to the input conditions during the training process. Experimental results demonstrate that the conditional feature concatenation mechanism avoids the interference of normalized conditions on timestep features inherent in traditional timestep embedding addition methods. Regulated by the joint random dropout strat-egy, the model can adaptively accept any number and combination of conditional inputs during the inference stage, outputting aircraft shape parameters that satisfy the corresponding set of constraints. This generative design model exhibits the capability of probabilistic inference under incomplete constraint inputs, providing a viable new ap-proach for aircraft conceptual design and the rapid exploration of aerodynamic layouts in multi-constraint scenarios.

Key words: Keywords: Multi-constraint design, Generative design, Diffusion Transformer, Fourier feature embedder, Conditional random dropout