Physics Informed Machine Learning for Accelerated Ship Hydrodynamic Optimization

The use of advanced computational fluid dynamics (CFD) methods, to calculate ship resistance during the design phase has been common practice for many years. However, these methods require many hours to perform a single calculation, creating a bottleneck in the hull optimization process. This limits the practical design space that can be explored to obtain the best design. On the other hand, data-driven machine learning techniques are well proven in other areas, but to replicate CFD would require massive amounts of data and training time, then work reliably only when interpolating between the known data. This project will explore the emerging field of physics-informed machine learning, combining both raw CFD data with known physical laws. The goal being to develop a faster method of predicting ship resistance, possibly 100 times faster than possible today, with comparable accuracy. This will significantly improve the number of designs that can be evaluated during the hull optimization process, leading to even more energy efficient ships in the future.

Today due to the time required to perform RANS CFD simulations, the ability to perform optimization is severely restricted. To address this problem this project seeks to explore the possibilities of state-of-the-art machine learning techniques and utilize them to accelerate the ship design optimization problem. It will review state-of-the-art techniques in classical data-driven machine learning. Then seek to improve upon these by adopting promising hybrid approach of physics-informed machine learning, which allows a model to be trained on a combination of raw data and knowledge of the governing laws of physics, like the Reynolds Averaged Navier Stokes equations. To perform the comparison of the various methods open CFD datasets will be produced, including parametric model to study hull variations on NTNU's research vessel RV "Gunnerus".