Towards Robust process design. The sensitivity analysis using machine learning methods
Résumé
Nowadays, mathematical models that describe the systems have become complicated dramatically and consist of more and more model parameters and input, which makes the uncertainty of model output increase. Accordingly, these implemented models in computer simulators are one of the main fundamental tools for understanding complex real-world. Sensitivity analysis (SA) and Machine learning (ML), as a part of these methods, have shown very satisfactory results with a wide range of applications, and all engineering fields are concerned, as far as they all need modeling and simulations. Sensitivity analysis is a certified method that is widely used to realize the model behavior and distinguish each input uncertainty's effect on the model output variability. It classifies into two main groups named Local SA (LSA) and Global SA (GSA). There are a wide variety of SA methods in the literature. The most commonly used methods of GSA are standardized regression coefficients (SRCs) and Sobol sensitivity indices. Both methods apply Monte Carlo simulations (MCS). However, this technique’s remarkable efficiency disadvantage is the high computational cost that makes it impractical for complicated systems. Moreover, ML and SA evaluate and work on the same problem from different visions. In general, the ML supervised approach aims to create a function that plots variables in the input space to variables in the output space. Most of the time, these functions are entirely data-driven. As noted before, SA focuses on the relationship between inputs and outputs. Each of SA and ML has some defects. As a result, SA and ML could provide considerable abilities for each method to profit from the other (Figure 1). For example, a critical weakness that ML is facing is the lack of interpretability and explainability. Also, surrogate modeling with the help of ML could help deal with the mentioned computational cost of the MCS in the context of SA. In chemical engineering and during classical process design, taking inaccuracies and uncertainties into account on the data is mainly carried out either by local SA (one parameter at a time) or by calculating flexibility indices. The global sensitivity analysis methods have employed more in numerical simulation in recent years, allowing uncertainty propagation modeling. Similarly, Machine Learning methods have been established with considerable expansion in dealing with classification, prediction, etc.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |