DOI
http://doi.org/10.1016/j.eng.2019.08.005

Summary
An idea for developing mesoscience-based artificial intelligence (AI) models is presented in this article. Mesoscience includes multiscale and multilevel analysis and can contribute to interpretability of deep learning. Firstly, the authors describe the evolution of AI and its wide scope of applications. They state that the popularity of AI was facilitated by the developed big data processing techniques, hardware technologies, and experts’ skills and competencies to apply AI to model large-scale complex problems. Secondly, the lack of interpretability of deep learning models is discussed. As many of the models are considered as ‘black box’ and only few of them can be explained for the specific tasks, the lack of interpretability is considered a bottleneck for the further development of AI. Finally, the strategy to integrate mesoscience in creating AI models is presented. The mesoscience approach is applied to improve model architecture, learning algorithms, computational methods. It is concluded that this approach is a promising strategy to improve interpretability of AI models.

Relevance for Complex Systems Knowledge
The authors suggest analyzing complex systems using multiscale approach which is usually applied in the chemical engineering. The main idea of this approach is to find relationships between the models in different scales as the direct application of artificial intelligence (AI) as a “black box” does not always represent the physical model. The authors suggest applying these principles to construct a deep neural network which is later used to solve the specific problem. They state that such principles better reflect human brain as people employ their understanding of physical nature together with the reasoning and assessment of abilities to obtain the correct solution.
Another level of complexity which is discussed in this article is complexity of big data. It is also related to the complexity of the model and is necessary to identify the physical mechanism of the complex system and make the created model effective.

Point of Strength
The point of the strength of this article is a suggested idea how complex systems can be analyzed using artificial intelligence and mesoscale approach which incorporates information about the system in different scales. Moreover, the complexity of big data is also discussed.