by A cutting-edge deep learning modeling technique that integrates physical knowledge is currently gaining popularity, with a variety of innovative methods emerging in the field. Physics-informed neural networks (PINNs) have been recognized as a prominent example of such techniques.
PINNs incorporate the residuals of the system’s governing partial differential equations (PDEs) and the initial value/boundary conditions into the loss function, ensuring that the resulting model adheres to the constraints set by the physical laws represented by the PDEs. However, PINNs face limitations in cases where equations defining the relationships among key physical quantities have not been established, necessitating the development of new approaches to model such systems.
A new deep learning modeling method known as Mesoscience-Guided Deep Learning (MGDL) has been introduced by Li Guo and a team from the Institute of Process Engineering (IPE) at the Chinese Academy of Sciences (CAS). The research paper detailing MGDL has been published in the journal Engineering.
Mesoscience is a methodology aimed at addressing multilevel complexities by studying mesoscale issues at different levels and linking macro-scale behavior and intrinsic system mechanisms through the principle of compromise in competition (CIC) among dominant mechanisms.
Unlike conventional deep learning methods, MGDL employs a unique approach to handling the dominant mechanisms of complex systems and their interactions based on the CIC principle of mesoscience when constructing sample datasets from system evolution data.
Mesoscience constraints are then incorporated into the loss function to guide the deep learning training process.
Two methods have been proposed to introduce mesoscience constraints to the training process, either as a loss regularization term or through learning rate adjustments. MGDL enhances the physical interpretability of the model training process by providing guidance based on physical principles and constraints.
The effectiveness of MGDL was assessed using a bubbling fluidized bed modeling case and compared against traditional techniques. The results indicate that model training based on mesoscience constraints offers significant benefits in terms of convergence stability and prediction accuracy, even with a smaller training dataset. This approach shows potential for broad applications across various neural network configurations.
Proposed by researchers from IPE, CAS, MGDL represents a novel strategy and methodology for leveraging physical background information during deep learning model training. With the growing prominence and application of mesoscience, MGDL is poised to be widely adopted for modeling complex systems.
*Note:
1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it
