Incorporating Artificial Intelligence into the Workflow for Calibrating a Numerical Wavetank

Recent research shows that, in addition to the universal approximation function, neural networks are capable of accurately approximating the mapping, both linear and non-linear, of one space of continuous functions into another. We report on our experience of coupling a well known CFD code, OpenFOAM, to neural network modelling using TensorFlow on a HPC system in order to model tank transfer functions. The CFD stage is well suited to using multiple CPUs via MPI libraries while the machine learning stage is better suited to using GPUs via several Python packages and libraries. We report that these distinct steps can be joined into an efficient pipeline by using the file system effectively. Our results show that neural network modelling works well in the region where non-linear CFD theories are needed to model the dynamics of water waves.