A reward-directed diffusion framework for generative design

This study presents a generative optimization framework that builds on a fine-tuned diffusion model and reward-directed sampling to generate high-performance engineering designs. The framework employs a parametric representation of the design geometry to generate new parameter sets that correspond to designs with improved aerodynamic and hydrodynamic performance. A key advantage of the reward-directed approach is its suitability for scenarios in which performance metrics rely on costly engineering simulations or surrogate models (e.g. graph-based, ensemble models, or tree-based) are non-differentiable or prohibitively expensive to differentiate. By using entropy-regularized approach, this work introduces the iterative use of a soft value function within a Markov decision process framework to achieve reward-guided decoding in the diffusion model. The proposed approach distributes computational and memory costs across the training and inference phases by integrating soft value guidance during both phases, thereby facilitating the generation of high-reward designs, even beyond the training data. Empirical results indicate that this iterative reward-directed method substantially improves the diffusion model’s ability to generate samples with reduced resistance in three dimensional ship hull design and enhanced aerodynamic performance in airfoil design tasks. The proposed framework generates samples that extend beyond the training data distribution, resulting in a greater 25 percent reduction in resistance for ship design and over 10 percent improvement in the lift-to-drag ratio for the airfoil design. These improvements surpass those obtained by traditional black-box optimization methods. Successful integration of this model into the engineering design life cycle can enhance both designer productivity and overall design performance.