Hull Design using Deep Learning Technology 

Technology is transforming the way that companies approach tough optimisation challenges, and Steller Systems is incorporating new, technically advanced methods into early stage hull design in ways not seen in some marine sectors before. Deep learning AI tools can allow millions of possible hull shape combinations to be used to identify optimised solutions during early stage hull design. 

Digital computing has developed over the past couple of decades and has revolutionised the speed of design development. Three dimensional models often replace two dimensional drawings entirely and simulation software has replaced the need for physical models and associated testing programs, due to being cheaper, quicker and easier for large data sets to be produced allowing design engineers to reach an optimised solution. 

Computational power is no longer a barrier to vast numbers of simulations being run in parallel, allowing data to form the backbone of most optimisations and can be readily created within sensible timeframes. Long gone are the days of expensive and slow simulations, partly thanks to the era of cloud-based computing, multi-license software and pay-as-you-go IT solutions. 

How do you ever know you have the most efficient hull? 

With ship resistance being directly proportional to fuel burn and with increasing environmental and financial pressures, every percent reduction in resistance is significant – especially when considered through the lifetime of the ship. However, CAPEX is often considered to be king over OPEX, and engines that potentially cost around 10% of a total platform cost could be downsized due to resistance reductions and still meet the same top speed requirement. This is a luxury that most companies cannot afford to ignore, leading to a positive design spiral of less weight, reduced power for auxiliary systems and reduced cost.  A more efficient hull can ultimately lead to lower fuel burn, smaller engines or a balance of both. 

Hull optimisation is a tough task, especially if trying to iterate towards a solution manually. Resistance in calm water, resistance in waves, seakeeping, geometrical constraints and operational profile all play their part in deriving a combination of shape parameters that best suits the final design. Like most things, it’s always a trade-off. But what if there was a way for example to find the lowest resistance hullform for a given operating speed profile and for a given range of displacements and centres of gravity? 

The now common tool of choice for deriving calm water resistance accurately is Computational Fluid Dynamics (CFD). Similar to decades ago when most design houses adopted CAD software to replace their drawings boards, many engineering firms are now fluent with CFD packages. A CFD resistance run may take an hour or two to process a single hull shape at a particular speed and displacement, which is pretty impressive compared with the time and cost implications of commissioning a physical model to be towed down a physical tank to produce the same values to similar accuracy.  

An hour or two does still look like a long time when you consider looking at thousands of combinations of a hull shape – defined possibly by 10 to 20 parameters – as well as varying speed and displacement to get to your optimum shape combination. Fortunately, although a complex optimisation challenge, there are trends in the objective function as parameters are varied. It is not like trying to manually unlock a 20-digit padlock because the trends let you know if you are getting closer or further from the optimum. It is these trends that need to be understood mathematically to allow the optimum results to be identified.  

What are Deep Learning Surrogates? 

Deep Learning Surrogates (DLS) are just one Artificial Intelligence (AI) tool that can be used to understand the trends and is already increasing the speed and effectiveness of automated design optimisation. The approach is based on  AI models  that have been key to solving other difficult computing problems, from tracking trends in the stock market to image recognition. 

Using DLS in the optimisation process looks very similar to traditional optimisation methods at the start. The constraints and the performance profile of the baseline design need to be clearly defined to bound the scope of the final solution and ensure it meets the same requirements as the baseline. Multiple simulations are then triggered with varying parameters to explore different design options and populate the design space. The similarities in the processes end about here. 

As the data from the simulations is produced it is used to train neural networks, and these are set up to effectively replace the original simulation tool. With sufficient data these deep learning models can predict an outcome significantly faster than the original simulation tool. In practice where it might take an hour to run another single simulation, the deep learning tool could predict a single result in seconds.  

Speed of predictions is where the real power can start to be realised, so long as the model has been fed with enough data to be accurate. Millions of interrogations can be performed in a short space of time to find the exact optimum within the design space. 

Methodology in how the design space is bounded becomes the critical priority to make the whole process as efficient as possible. There is no point in looking at an area that is known to be poor or is not within the constraints as it can waste effort, however generating designs that do not quite meet the constraints can help generate data around the edge of the design space highlighting the constraint that is driving the ultimate performance.  

DLS for hull optimisation 

In hull design, minimum stability is an example of a dominant constraint that can drive the final outcome. Beam is a powerful influence on both resistance and stability so setting the minimum and maximum range of beam to be explored will affect the effort required to achieve the outcome. Too little beam in the simulation data used to feed the model will generate a lot of data for unstable hulls, and too much beam will most likely yield high resistance, but very stable contender solutions. 

By exploring the entire design space and generating millions of combinations of hull shapes it stands to reason how significant savings can be realised. A recent task to optimise a future frigate hull form showed the extent of the potential savings. A 32% reduction in resistance at cruise speed (16 knots) and for the same hull 22% reduction in resistance at top speed (24 knots) was achieved. Putting this into perspective, this allows the designers to save 5 megawatts from a 30 megawatt drive train, showing how large the savings can be. 

Novel applications of deep learning in engineering development, and more specifically in ship design, is only going to increase. Steller Systems is also developing a tool to optimise damaged stability in naval vessels where survivability is more critical compared with the commercial marine world. Already this is proving that it is possible to match survivability, with less subdivision and hence reduced cost. Propulsor and structural optimisation tools are also being developed. More information on this to follow. 

What has become clear to us at Steller Systems is that before considering novel, costly, higher maintenance systems to reduce your carbon footprint or increase performance, optimising the design of the hull shape, subdivision, propulsor and structure should be the first priority above all else.  

Find out more about HullTune® by Steller Systems. If you are interested in how we can use Deep Learning Surrogates to help your project, then please get in touch.