Key Findings
Breakthrough technology combines traditional physics-based modelling with artificial intelligence to improve oil spill trajectory predictions significantly.
Significant accuracy gains achieve up to 20% more precision in matching satellite observations and 25% more accuracy in tracking oil slick position.
Real-world validation successfully demonstrated on the 2021 Baniyas oil spill in Syria, where over 12,000 cubic meters of oil entered the Mediterranean
A transferable framework that can be adapted to other environmental forecasting systems facing similar uncertainty challenges.
Oil spills can be among the most devastating environmental disasters, with the potential to severely damage marine ecosystems, disrupt coastal communities, and impose lasting economic damage. Traditional numerical models, such as MEDSLIK-II , simulate the movement and transformation of oil particles in seawater, but their accuracy has been limited by dependence on expert judgment for tuning critical physical parameters. This manual calibration process, while informed by experience, is not always able to capture the complexity and variability of real-world ocean and atmospheric conditions.
The new study, published in the journal Ecological Informatics, " Improving oil slick trajectory simulations with Bayesian optimization ", addresses this challenge by introducing Bayesian optimization – an artificial intelligence (AI) technique that automatically learns from satellite observations to adjust the model's physical parameters. A hybrid approach that is able to combine the reliability of physics-based modelling with the efficiency and adaptability of AI.
"This work represents a significant step forward in narrowing the gap between traditional numerical ocean modelling and AI methodologies, showing that hybrid solutions can effectively harness the strengths of both worlds," explains CMCC and Columbia University researcher, and lead author of the study Gabriele Accarino . "By coupling the Bayesian optimization framework with the widely used community MEDSLIK-II oil spill model and satellite-based observations, we have introduced a prototype for next-generation operational forecasting systems."
Real-world applications
"Oil spills have serious impacts on ecosystems and human activities, and predicting their evolution is crucial for effective interventions," notes CMCC researcher and co-author of the study Marco De Carlo . "Traditional numerical models are useful but rely on manually chosen parameters, which can introduce uncertainty. Rather than replacing physics, our hybrid approach complements it, enhancing the realism and reliability of simulations, and performing well even with sparse data."
The effectiveness of the hybrid approach was validated by the research team, using the 2021 Baniyas oil spill incident in Syria , where more than 12,000 cubic meters of oil entered the Mediterranean Sea. The results obtained in this application demonstrated remarkable improvements in forecasting accuracy: spatial accuracy increased by up to 20% in matching satellite observations of the oil slick's shape and spread; position tracking improved by up to 25% compared to standard model predictions; and the overall skill score (which compares spatial distributions of oil spill against a ground truth such as satellite observations) improved from 7.97% to 20.66% on average compared to control simulations.
These improvements were consistent across multiple time steps, particularly during periods of increased drift variability, demonstrating the method's effectiveness in dynamic environmental conditions. This comes with significant benefits for emergency response services during oil spill incidents, as more accurate trajectory forecasts enable authorities to deploy response efforts more effectively, potentially preventing further damage to marine ecosystems.
Another important advantage is that the trained machine learning model helps speed up the numerical model by efficiently calibrating its parameters, allowing faster analyses and scenario testing.
"It can also update in real time as new observations arrive, and the framework is transferable and relocatable, allowing it to be applied to different geographic areas or other contexts, such as atmospheric or oceanic modeling," continues De Carlo. "This makes it not only a research tool, but also a practical solution for operational uses, supporting rapid decision-making during environmental emergencies."
"As an oceanographer specializing in marine pollution simulations, I know that the modeller's experience is crucial for the successful representation of events such as oil spill leakages," says CMCC researcher and co-author of the study Igor Ruiz Atake . "The more one knows about the area of interest and the simulation tool, the better the results will be. Our interdisciplinary team at CMCC has developed this new approach that automatically searches for the optimal tuning parameters of the oil spill model. The results we obtained still need to be tested on other real spill events. However, from our findings, we expect it to save time that experts can use to gain a deeper understanding of the event as a whole, instead of spending time on the technicalities of the problem. In marine emergencies, time is of the essence. "
Beyond oil spill response
This innovative framework also offers significant potential for adaptation in other environmental forecasting systems that face similar challenges of uncertainty and limited observations. For example, the approach could also be applied in atmospheric and general ocean circulation models, potentially reducing long-standing modeling biases and improving the representation of small-scale physical processes.
"In this sense, the study not only introduces a novel technical contribution but also points toward a paradigm shift in environmental forecasting, where physics-informed AI becomes a cornerstone of operational risk management and climate resilience strategies," says Accarino.
As climate change continues to alter ocean and atmospheric conditions, innovative approaches like this AI-enhanced modeling system become increasingly crucial for protecting marine environments and coastal communities from environmental disasters.
The study's findings demonstrate that the integration of artificial intelligence with traditional environmental modeling can deliver practical improvements that benefit both scientific understanding and real-world emergency response capabilities.
The simulations used in the study were conducted with CMCC's JUNO Hybrid Cluster , one of Europe's most advanced computing facilities for climate and environmental research. CMCC researchers led the design of the optimization workflow, integrated the machine learning components, and validated the results against satellite data.
