Qualifying and Regulating the Use of Artificial Intelligence in Software Systems for the Canadian Nuclear Sector

Abstract

This thesis examines the inadequacies in qualifying Artificial Intelligence (AI) software for the Canadian nuclear energy sector. The nuclear energy sector is a high-risk environment with strict regulations to ensure safety. Despite the rising popularity of new technologies like AI, a compliance assessment would be needed against nuclear qualification procedures. First, the thesis analyzes the existing regulatory framework within the Canadian nuclear sector. This analysis reveals potential gaps that traditional software qualification methods fail to address when applied to AI. The risks of AI, primarily linked to the complexity and opacity of decision-making processes, show the need for a new approach to AI regulation in nuclear. Next, a review of the Canadian regulatory framework focusing on the Canadian Standards Association (CSA) N290.14 with case studies of qualifying commercial software is presented to showcase the software qualification lifecycle. Through these detailed case studies, gaps are identified when applying the software qualifications methods to AI software. The thesis also suggests three methodologies for future AI qualification: model interpretability, feature importance, and data variety. These features are investigated in order to improve the transparency, reliability, and safety of AI applications in high-risk contexts such as nuclear power plants. Finally, the thesis proposes incorporating these three methodologies of AI into the software qualification framework to significantly mitigate the risks and support a safe deployment and operations of AI based software in the nuclear sector.