TLDR: A new method called “Censored Sampling” improves AI-generated structural designs by incorporating human feedback. It trains small reward models from human evaluations to guide diffusion models, reducing common design flaws like disconnected parts or boundary issues without needing to retrain the main AI model. This makes AI-designed structures more realistic and manufacturable.
Designing strong and efficient structures, from airplane parts to bridges, is a complex task known as topology optimization. Traditionally, engineers used methods that were very time-consuming and often produced designs that were hard to manufacture. In recent years, artificial intelligence, particularly machine learning, has offered faster ways to generate these designs.
One powerful AI technique is the use of denoising diffusion models. These models, like the TopoDiff framework, can create optimized structures by learning from existing designs. However, even these advanced AI models sometimes miss subtle but critical design flaws. For example, they might generate structures with parts that float unconnected or don’t properly attach to their supports. While these issues might not trigger automated error checks, they are immediately obvious to human experts.
This is where a new approach, called Censored Sampling for Topology Design, comes in. Researchers Euihyun Kim, Keun Park, and Yeoneung Kim have developed a novel framework that brings human judgment directly into the AI design process. Their method guides the AI’s generative process using a lightweight ‘reward model’ that learns from minimal human feedback. This means the AI learns to avoid creating unrealistic or unmanufacturable designs by understanding what human experts consider flawed.
The core idea is inspired by how large language models are aligned with human preferences. In this system, human experts provide simple ‘yes’ or ‘no’ feedback on AI-generated designs. They might indicate if a structure has a ‘boundary condition violation’ (not properly connected) or ‘floating material’ (disconnected parts). This binary feedback is then used to train small, specialized AI models called reward models.
These reward models don’t retrain the main diffusion model from scratch. Instead, they act as a guide during the design generation process. As the AI creates a structure, the reward models provide signals that steer it away from designs with the identified flaws. It’s like having a human expert constantly whispering corrections to the AI as it draws, ensuring the final output is not only structurally sound but also physically plausible and manufacturable.
The researchers conducted several case studies to demonstrate the effectiveness of their method. They showed how designs that initially had broken boundary conditions or floating material were corrected when the human-guided sampling was applied. Even in cases where designs had multiple types of flaws, the combined guidance from different reward models successfully resolved them. Quantitative evaluations further confirmed that this human-in-the-loop approach significantly reduces the rate of invalid designs compared to the baseline AI model.
A notable aspect of this work is its modularity. The human feedback system can be easily integrated into existing diffusion models without needing extensive retraining. Furthermore, the process can be refined iteratively; new AI-generated designs can be presented to humans for further feedback, continuously improving the reward models and the quality of the outputs.
Also Read:
- Guiding Diffusion Models for More Consistent Time Series Forecasts
- New Sampling Method Boosts Efficiency in Predicting Structural Failures
While this method marks a significant step forward, it does have some limitations. The quality of the designs depends on the quality and diversity of the human feedback. Also, the current system uses simple binary labels, which might not capture all nuances of design quality. However, this research paves the way for more reliable and trustworthy generative design systems, effectively bridging the gap between automated design generation and expert judgment. For more details, you can read the full research paper here.
