An interpretable deep learning modeling architecture considering process underlying logics reveals a promising way to the intelligent chemical industry

Intelligent manufacturing is one of the most important strategies towards the upgradation of chemical industry. The deep learning modeling technology acts as a crucial role in this industry upgradation with its strong fitting and predicting ability. At the same time, such intelligent modeling technology is still challenged by the complexity of chemical processes, leading to limited generalization performances in practice. Therefore, the chemical industry calls for more interpretable and generalized deep learning modeling architectures.

In a recent research paper published in the journal of Engineering, Prof. Weifeng Shen’s team at Chongqing University presents a novel interpretable deep learning modeling architecture for chemical processes, which is called light attention–convolution–gate recurrent unit (LACG). To deal with the complex variable interactions in chemical processes, they classify the driving forces of the process variable fluctuation into three parts: the general dynamic behavior, transient disturbances, and other input factors. As such, the new modeling architecture adopts three different deep learning modules to learn the three driving forces, respectively, and the high-accuracy modeling result on dynamic chemical processes can be obtained via the elaborate coupling of three sub-modules based on underlying logics.

Differing from widely used deep neural networks such as the feedforward neural network, convolution neural network, long-short term memory (LSTM), the deep learning modeling architecture proposed in this work is designed according to the characteristics of actual process dynamic behaviors. More specifically, convolution layers are adopted to fit the temporal interactions of process variables, a light attention module is proposed to capture the instantaneous variable fluctuations, and a residue module including gate recurrent unit (GRU) is embedded to eliminate interference factors along time.

Considering the process underlying logics, the proposed modeling architecture delivers a high-fidelity modeling result on an industrial deethanization process, and shows its robustness on various datasets. Meanwhile, the authors make an elaborate study on the interpretability of the built deep learning model. The parameters extracted from the built model are proved to be interpretable that the process dynamic behaviors revealed by such parameters are consistent with the simulated results of the deethanization theoretical model. It is a novel experiment to explore the interpretability of deep learning model by comparing with the theoretical simulated results. In addition, comparison is made between the interpretabilities of the proposed modeling architecture and the widely used deep neural network, attention-based LSTM. The new architecture obviously outperforms the attention-based LSTM in terms of interpretability, with more details on the process variable interactions.

In conclusion, this research provides a practical deep learning modeling architecture for chemical dynamic processes. Compared with traditional deep neural networks, the clear advantages of the proposed modeling architecture in terms of prediction accuracy, generalization, and interpretability make it an efficient and reliable approach towards the intelligent system construction in future chemical industry. Intelligent manufacturing is an inevitable route for chemical industry facing with various challenges such as dual carbon targets and sustainable development. This research inspires the studies on the intelligent modeling of chemical processes by completely integrating artificial intelligence technologies and chemical process principles. It helps to solid the theoretical basis and application foundation for intelligent chemical engineering, and promotes the development of new quality productivity.

The paper “An Interpretable Light Attention–Convolution–Gate Recurrent Unit Architecture for the Highly Accurate Modeling of Actual Chemical Dynamic Processes,” authored by Yue Li, Ning Li, Jingzheng Ren, Weifeng Shen. Full text of the open access paper: https://doi.org/10.1016/j.eng.2024.07.009. For more information about the Engineering, follow us on Twitter (https://twitter.com/EngineeringJrnl) & like us on Facebook (https://www.facebook.com/EngineeringJrnl).