Reproduce “Physics-informed Diffusion Generative Model for Time-Series Data Synthesis in Dynamic Systems” on OpenResearch

About this paper

Haiteng Wang, Yunfei Zhu, Tao Wang, Yikang Li, Jiabao Dong, Xiaoge Zhang, Lei Ren

PhysDGM introduces a physics-embedded diffusion generative model that synthesizes high-fidelity, physically consistent time-series data for dynamic systems by integrating physical laws throughout its iterative generation process. The model significantly enhanced predictive utility across industrial AI tasks and achieved performance comparable to full real datasets using only 5% of original training data.

Key insights

• Enforcing physical constraints iteratively at each step of a diffusion model is more effective than applying them only at the final output, leading to superior physical consistency in generated time-series trajectories.
• Gradually intensifying physical constraints during model training, rather than imposing them rigidly from the start, optimizes the learning process by allowing the model to first capture the data distribution before refining for physical adherence.
• Leveraging gradient guidance from new physical constraints during the sampling process allows for zero-shot adaptation of the generative model to evolving physical conditions without the need for computationally expensive retraining.

Summary by alphaXiv.