Polynomial Neural Networks for improved AI transparency: An analysis of their inherent explainability (operational rationale) capabilities

Abstract

The demand for reliable Artificial Intelligence (AI) models within critical domains such as Chemical Engineering has garnered significant attention towards the use and development of transparent AI methodologies. Nevertheless, the field of AI transparency has received an uneven level of attention, such that crucial aspects like explainability (i.e., the transparency of the AI's operational rationales) have remained understudied. To address this challenge, this study investigates the inherent explainability capabilities of Polynomial Neural Networks (PNNs) for applications within Chemical Engineering. PNNs, which implement higher-order polynomials in lieu of linear expressions within their hidden layer neurons, are inherently nonlinear, and thus do not require an activation function to accurately capture the behavior of a system. Accordingly, these neural networks provide continuous, closed-form algebraic expressions that can be used to ascertain the contributions of individual features in the AI architecture towards the network operational behavior. In order to study this behavior, the PNN method was adopted in this work to capture the relationships of noiseless and noisy data derived according to simple mathematical expressions. The PNN polynomials were then extracted and examined to highlight the insights they provide regarding the system operational rationales. The PNN method was furthermore applied to capture the behavior of a circulating fluidized bed reactor to fully showcase the explainative capability of this method within a Chemical Engineering application. These studies highlight the intrinsic explainability capabilities of PNNs and demonstrated their potential for reliable AI implementations for applications in Chemical Engineering.