Multi-source eXplainable Artificial Intelligence for monitoring and interpretation of marine scrubber washwater quality from full-scale vessel data.

Marine Scrubbers enable ships to achieve full compliance with sulphur emission limits by effectively removing SOX from exhaust gases, in line with the requirements set by the International Maritime Organisation (IMO) under MARPOL Annex VI, considering both gas and washwater quality standards. This technology represents a reliable and widely adopted solution for reducing atmospheric emissions. However, its operation generates washwater streams that require careful monitoring. Key parameters, including polycyclic aromatic hydrocarbons, turbidity, and pH, are affected by interacting operational, fuel-related, combustion-related, and seawater conditions. In this work, a multi-source eXplainable Artificial Intelligence (XAI) framework was developed to predict and interpret scrubber washwater quality using real-world data from two full-scale vessels. Continuous emission monitoring system data, bunker notes, voyage information, and environmental descriptors retrieved from the Copernicus Marine Service database were integrated into a unified database, and five multi-output regression algorithms were benchmarked for the prediction of outlet PAH, outlet turbidity, and outlet pH. The Random Forest algorithm achieved the best overall performance, with test-set R2 values of 0.993 for PAH, 0.894 for turbidity, and 0.997 for pH. A hierarchical explainability workflow based on permutation feature importance, SHAP, PDP/ICE, and LIME was then used to identify dominant drivers, target-specific contribution patterns, and response-shape behaviour. The results show that the model is mainly governed by variables related to gas-side pollutant loading, fuel sulphur, combustion regime, hydraulic operating conditions, and inlet turbidity. In particular, turbidity-related information emerged as an interpretable and operationally meaningful indicator candidate for PAH dynamics. The study demonstrates that XAI can support not only accurate prediction, but also physically grounded insight for more transparent marine scrubber washwater monitoring.