Software: Stochastic Prediction of Oil Spill Transport and Fate using Ap-proximation Methods or Machine Learning

Abstract

Oil spills represent a persistent risk to marine ecosystems, coastal communities, and energy-related maritime operations, demanding predictive tools that areboth accurate and computationally efficient for real-time decision support. This paper presents CRANSLIK 3.2, a machine-learning-driven oil spill trajectoryforecasting system for the Mediterranean Sea that significantly advances CRANSLIK 3.1's capabilities through targeted optimisation and architecturalrefinement of deep neural networks. The system is validated against the established MEDSLIK-II model using a documented real-world oil spill event off thecoast of Algeria, demonstrating reliable operational performance. Key innovations include the application of the Levenberg–Marquardt optimisation al-gorithm and a comprehensive evaluation of Long Short-Term Memory (LSTM) network architectures, in which fourteen activation function combinations weresystematically tested. An LSTM configuration combining Exponential Linear Unit (ELU) activation with Sigmoid gating functions achieved the highestpredictive accuracy while pre-serving rapid inference times. The results highlight the ability of data-driven models to complement physics-based approaches,offering a robust, scalable, and time-critical forecasting tool for environmental protection and energy-sector risk mitigation.