Scientific machine learning for modeling industrial-scale Primary Separation Vessel

A scientific machine learning (SciML) method integrates masked neural networks with fundamental physical laws to model the Primary Separation Vessel (PSV) in oil sands processing. This framework addresses limitations in conventional models by discovering a critical underlying relation, namely optimized hindered settling functions, through embedded neural networks, subsequently translated into interpretable mathematical expressions via symbolic regression. The methodology significantly enhances computational efficiency through parallel processing and pseudo-transient solver techniques while eliminating the need for scenario-specific parameter tuning. The re-discovered hindered-settling function successfully captures the fundamental behavior of hindered settling across varied conditions, although parameter estimation relied on a limited dataset. Validation against industrial range benchmarks demonstrates the approach’s effectiveness in reproducing characteristic physical behaviors across different ore grades. This advancement represents a significant step in bridging theoretical and data-driven approaches for complex multiphase systems, with promising potential for extension to integrated systems and improved operational optimization in oil sands processing facilities.