.png){kind=icon}
Return
Dynamic SHAP: Time resolved explainable AI for batch processes
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compchemeng.2026.109717.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the field of chemical engineering, data-driven models offer an appealing alternative to first-principles models, as they do not require a detailed quantitative description of the physicochemical phenomena involved. This is particularly relevant in batch processes, where in many cases the intrinsic dynamic operational nature and high chemical complexity make physics-based models difficult to apply to real-world problems. While in the past the majority of machine learning (ML) models applied to batch processes were linear, nowadays, the increase in availability of computational power allows exploring the use of more complex ML algorithms. However, nonlinear ML methods might lack transparency and interpretability compared to their linear counterparts. This study introduces a novel explainable Artificial Intelligence (AI) analysis method specifically tailored for batch process analysis: Dynamic SHAP (D-SHAP). D-SHAP extends the well-known SHAP (Shapley Additive Explanations) analysis to batch data analysis. The D-SHAP framework transcends simple feature aggregation by reconstructing continuous importance trajectories, enabling the identification of critical temporal windows and aligning model interpretability with the phase-dependent dynamics of batch processes. Therefore, it allows pinpointing not only the most influential variables but also the specific phases and time within a batch where each variable exerts maximum influence. The proposed explainable AI analysis is demonstrated on three use cases and three ML methods: (i) Partial Least Squares, (ii) Boosted Tree and (iii) Random Forest.
Keywords:
Machine learning
Batch processes
Explainable AI
D-SHAP
Journal
C
IF:
3.9
Papers:
279
Citations:
0
Organization
Cited Papers
No cited papers available