How evolutionary AI can help save fuel for ships

Haoyu Yang

How are modern transportation vessels different from the older generations, and how can modern machine learning technologies help those giant sailing machines reduce costs? With these questions I started my research early this year. The results are remarkable and promising.

Shipping goods across the oceans can date back to ancient times. The field of artificial intelligence (AI) research was founded as an academic discipline in 1956. However, it is the recent development from both fields that enable AI technologies to be of help for ship operations. Implementing the evolutionary AI tool, I found a way how we can help client shipping companies save fuel costs with our data driven solutions.

Development of ship operation data

Noon reports have been the dominating way to collect and report ship operation data for the past several decades ever since the invention and prevalence of telex and radio. A noon report is a daily data sheet prepared by the ship’s chief engineer. The report provides the vessel’s position and other relevant standardised data to assess the performance of the ship based on its speed and environmental forces including weather conditions.

The example form below shows what a typical noon report looks like. There definitely are more complicated noon reports, but it is safe to say that such a daily report barely collects more than 100 data points per day. Note to say that the daily aggregated data does not give any insight regarding the data variation throughout the whole day.

img1Example of a noon report

With the advancement of ship onboard machineries and IoT technology, there are growing numbers of sensors recording hundreds of data points about ship operations every second. Enabled by the modern communication technologies, all these millions of daily data points - or at an aggregated level - can be send back to the control centers on shore in real time. The table below is an example I have worked with during my internship at Mirabeau, a Cognizant Digital Business. There are almost 500,000 data points collected by sensors on board of a ro-ro transportation vessel (ro-ro means roll-on/roll-off, ships that carry wheeled cargo, such as cars, trucks, semi-trailer trucks, trailers, and railroad cars) on March 5th 2020, only upon the 41 measurements I chose for my project.

img2Data points collected by sensors on board a ro-ro transportation vessel in one day

This vast amount of data enables analytic technologies to help with decision making at all levels:

  • Ship operation analysis (quantitative based) can be done with data from the past months or even weeks instead of past years
  • Real-time operation optimization onboard
  • Real-time operation optimization onshore
  • Fleet management and control based on data driven decisions
  • Voyage planning based on more detailed historical data

As for my project, I choose to utilize the sensor data for onboard real-time trim optimization to save fuel for the ship. Only data collected at very frequent interval in real time onboard of a modern ship would allow for this kind of optimization.


The trim of a ship is the difference between the forward and backward draft. The draft or draught of a ship's hull is the vertical distance between the waterline and the bottom of the hull (keel), with the thickness of the hull included; in the case of not being included the draft outline would be obtained. Draft determines the minimum depth of water a ship or boat can safely navigate. The figure below shows the definition of trim.

img3Trim of a ship

Evolutionary AI (LEAF-ESP)

The tool employed to achieve this real-time optimization is one of the functions of the evolutionary AI tool from Cognizant, Learning Evolutionary AI Framework (LEAF) – Evolutionary Surrogate-assisted Prescriptions (ESP). ESP consists of two models. The first model is the Predictor that can be any machine learning model, such as a random forest or a neural network. It is trained on historical data to map contexts and actions to outcomes. The second model is called the Prescriptor, which is a neural network that is constructed through evolution. The Prescriptor maps the contexts to actions that lead to desirable (optimized) Outcomes.

In case of my project, the desired outcomes of the model are a pair of variables with an interesting trade-off, fuel consumption and speed. Essentially for a sailing vessel, we would like to achieve the best speed (required) with the least fuel consumed. The Predictor was trained first with a static dataset collected by ship sensors of 62 days and then the Prescriptor evolved against it and prescribes the optimal trim for the ship at different timestamps from the testing set. Once the system is embedded in an outer loop with the real world, where the prescriptions are implemented, more training data can be obtained for the Predictor. Following this loop, ESP will recommend more accurate and efficient actions overtime. The more and longer the model is in use, the better value it has. The picture below depicts such a loop.

img4LEAF-ESP lifecycle: continuous learning and optimization

How exactly will the optimal trim prescribed by the model be presented to the captain and onboard crew for them to make the actual decisions? A dashboard will help to keep an eye on the sensors and equipment on board.

A dashboard could constantly monitor the vessel and can highlight where there are issues with equipment connected to the system. Systems with abnormal data observed will show a red alert as a reminder to investigate. More importantly, the on-board console provides live-on current trim status and the optimal trim for the current sailing conditions and the speed required. This allows the crew to make informed decisions to optimize the vessel’s performance and reduce fuel accordingly.

According to my research, the optimal trim values prescribed by the trim optimizer results in 15% saving on fuel when tested against the test set of data.

Future thoughts

Besides the direction of trim optimization, there are much more other directions on how to utilize the vessel IoT sensor data with the help of LEAF. Listed below are a few interesting directions of great potential in my view:

  • Ship health and maintenance advices
  • Routing planning for a voyage
  • Speed planning for different parts of a voyage
  • Higher level fleet management among different ships

Will dashboards and predictive models change the course of future shipping and management of shipping fleets? In any case, I am proud to have contributed in this field of research and I am excited to see the increasing maturity of AI application and IoT data usage in shipping industry. I predict a better-informed and greener ship fleets, powered by data driven methods.