DDIM-based hybrid accelerated sampling method for generative design of aircraft configuration

Traditional aircraft configuration design faces significant efficiency challenges. This study addresses the critical bottleneck of slow sampling in point cloud diffusion models for generating 3D aerodynamic configurations under multidisciplinary constraints. We introduce the denoising diffusion implicit model (DDIM) acceleration strategy to substantially reduce the required sampling iterations, leveraging its non-Markovian skip-step mechanism without model retraining. Specifically, reducing the sampling steps from 1000 to 50 cuts the generation time by 57.5% (from 30.32 s to 12.89 s), while the average aerodynamic performance relative error increases only from 3.62% to 8.52%. To further optimize the balance between speed, accuracy, and diversity, we propose a novel “deterministic-stochastic” hybrid sampling strategy. This approach dynamically identifies critical timesteps by analyzing the temporal evolution of latent point cloud feature gradients and employs a trained classifier to adaptively modulate the noise strength parameter (η) across regions of varying criticality. Experimental validation demonstrates that the hybrid strategy operating at 50 steps delivers generation time below 15 s, achieves a 76.6% satisfaction rate for Coverage (COV, chamfer distance) below 10%, and attains a 73.3% satisfaction rate for aerodynamic performance error below 10%, outperforming static noise sampling. This work successfully integrates DDIM acceleration with dynamic noise regulation into a point cloud diffusion framework for aircraft configuration generation, effectively overcoming the sampling efficiency hurdle and enabling the rapid production of diverse, constraint-satisfying designs. Future efforts will focus on automating the optimization of the classifier and noise control parameters.