Generating mechanistic explanations of process systems requires identification of mechanistic process models, and relationships between the process variables. In order to infer the causal interactions between system variables requires tremendous human effort based on prior knowledge about the system. To create an automated system to generate such explanations of the system, the system needs to be equipped with plausible mechanisms, and scientific knowledge about physicochemical model forms, variables and parametric mapping.
In this talk, we present an artificial intelligence system, XAI-MEG, for generating natural language explanations of process systems from process data, using an ontology of physicochemical phenomenological interactions. These first-principles interactions are codified in an ontology of model forms, process variables and physical process parameters. The system is equipped with knowledge about physicochemical processes, variables, known feature transformations of the variables, scientific model forms, and explanations of the model form, based on simplifications of phenomenology and conservation laws. We show how ontology of symbolic variables and contexts can be combined with machine learning to create causal explanations of process systems.
