Long-short-term memory (LSTM) prediction of marine oil spreads

We assess the feasibility of utilizing deep learning, particularly the Long Short-Term Memory (LSTM) algorithm, for predicting the spreading of low-sulfur fuel oil (LSFO) under accidental marine spills. Training data was generated through numerical simulations of various artificial geometries, incorporating different configurations of islands and shorelines, as well as varying inlet wind speeds (2.0–8.0 m/s). We performed a three-phase simulation by assuming oil as a liquid at a scale of several hundred meters. Based on these results, we extended the scale to kilometers by assuming oil particles. For simulating the spread of oils over O(10²) km scales, the volume of fluid and discrete phase model were utilized. Key kinematic variables, such as particle location, particle velocity, and water velocity, were collected as input features for the LSTM model. The predicted LSFO spreading pattern showed a strong correlation with the simulation results, exhibiting less than 10% mean absolute error for the untrained data. The model was further validated by applying it to the actual Wakashio LSFO spill accident together with real geometric and weather data, thereby confirming the practical feasibility of the proposed model.