(259a) Process Safety and AI

Understanding process safety requires “big data” analysis involving artificial intelligence (AI). Fortunately (or unfortunately from a data-analysis perspective), expensive plant shutdowns and deadly accidents occur very rarely. And, consequently, new techniques are sorely needed to assist operators and plant managers better appreciate the need to take preventive actions to circumvent them. In this presentation, several approaches are reviewed, including dynamic risk analysis (DRA) using near-miss alarm data and Bayesian statistics – augmented with model-informed prior distributions as processes approach rare events. These yield improved posterior distributions that predict failure probabilities. Similar to model-predicate control (MPC), when safety systems come into play, new strategies for model-predictive safety (MPS) predict more reliably the inability to avoid rare events. Also, from molecular modeling (such as molecular dynamics), statistical model-driven, path-sampling techniques help identify rare paths that move from normal operation to rare-event basins of attraction. These include transition-path sampling and forward-flux sampling. Applications of these approaches to industrial processes are considered.