(287e) Explainable AI: Generating Causal Explanations of Machine Learning-Derived Models from Data | AIChE

(287e) Explainable AI: Generating Causal Explanations of Machine Learning-Derived Models from Data

Authors 

Sivaram, A. - Presenter, Columbia University
Venkatasubramanian, V., Columbia University
One of the main concerns of purely data-driven black-box models obtained via machine learning techniques is the lack of their explanatory power. On the other hand, mechanistic models derived from fundamental physiochemical mechanisms do not suffer this limitation, but they require considerable effort and expertise to develop. Here we present a hybrid system, an explainable AI-based modeling engine (X-AIME) that generates natural language-like mechanistic explanations of the models automatically discovered from data. Our philosophy is to combine certain aspects of expert systems, lessons learned from the largely forgotten era of symbolic AI, with the numeric AI of modern machine learning. This hybrid system is equipped with knowledge about fundamental physicochemical processes, variables, feature transformations of the variables, common model forms and their elementary explanations. We demonstrate the end-to-end information pipeline through which the system generates mechanistically feasible models and automatically generates explanations about the models using test cases from reaction engineering and transport phenomena.