A Simple Idea That Solves Complex Problems in System Modelling
Professor Yun Li's team built a grey-box model that combines physical laws with neural networks and found it modelled a cleanroom temperature control system better than conventional approaches.
Read the original paper
During efforts to improve precision engineering, a new method has surfaced to break the limitations of current modelling techniques. Researchers Dr. Chen Luo, Ao-Jin Li, Jiang Xiao, and Ming Li, led by Professor Yun Li from the Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, have introduced a practical solution. Their study, published in Scientific Reports, explains a method called grey-box state-space model (SSM), which combines simplicity, accuracy, and transparency for dynamic modelling.
Combining basic scientific principles with advanced data analytics, this grey-box hybrid model merges a white box of physical laws, which are symbolic rules that describe how things like motion and energy behave in the real world, and machine learning techniques with a number of black boxes using universal function approximators like connected artificial neural networks, which involve data-driven training or prediction. This combination creates a model that not only interprets but adjusts to varying complexities in real-world scenarios. “By incorporating expert knowledge within a strong AI framework, we ensure that these models are understandable and effective under different conditions,” said Professor Li.
Testing this approach on a highly sensitive temperature control system used in cleanrooms for manufacturing demonstrated its effectiveness. These systems, which are environments free from dust and contaminants, demand extremely accurate temperature regulation for both air and water. The grey-box model exceeded the performance of traditional methods, managing unpredictable system changes and unique characteristics better than standalone approaches.
Developed with an SSM structure, the grey-box model uses two transformations. One transforms an irregular nonlinear differential equation set into a regular, linear-like global SSM white box, and the other transforms its state-dependent parameters into regular local function approximators. Thus physical laws form the foundation of the model while employing machine learning to adjust parameter settings dynamically. For instance, in cleanroom air temperature control, this model relied on both energy transfer principles, which explain how heat moves between objects, and real-time data, which is information collected as events happen, to achieve optimal performance. “Our model can predict behavior in new scenarios with remarkable accuracy,” Professor Li explained, “making it essential for industries where operating conditions often change.”
Solving common challenges like incomplete information, which refers to gaps or missing data, and inefficient calculations, the grey-box framework displayed a higher ability to adapt while still providing insights into how it works. This blend of adaptability and clarity is essential for practical industrial use.
Future possibilities for grey-box SSM span diverse fields, including aerospace, which involves the design and production of aircraft and spacecraft, and energy management, which focuses on using resources efficiently. Professor Li sees this method as part of a broader move towards smarter, more transparent technology in engineering. This shift represents a future where machines not only perform but also explain their functions, enhancing both trust and efficiency. Professor Li remarked, “Our goal is to develop efficient and explainable ‘AI for Engineering’ tools.”
About the researchers
Journal reference
Luo, C., Li, A., Xiao, J., Li, M., & Li, Y. “Explainable and Generalizable AI-Driven Multiscale Informatics for Dynamic System Modelling.” Scientific Reports, 2024. https://doi.org/10.1038/s41598-024-67259-4







