Neural network as a function approximator and its application in solving differential equations

A neural network(NN) is a powerful tool for approximating bounded continuous functions in machine learning. The NN provides a framework for numerically solving ordinary differential equations(ODEs) and partial differential equations(PDEs)combined with the automatic differentiation(AD) technique. In this work, we explore the use of NN for the function approximation and propose a universal solver for ODEs and PDEs. The solver is tested for initial value problems and boundary value problems of ODEs, and the results exhibit high accuracy for not only the unknown functions but also their derivatives. The same strategy can be used to construct a PDE solver based on collocation points instead of a mesh, which is tested with the Burgers equation and the heat equation(i.e., the Laplace equation).

COMPUTATION OF STRESS INTENSITY FACTORS BY THE SUB-REGION MIXED FINITE ELEMENT METHOD OF LINES

Based on the sub-region generalized variationM principle, a sub-region mixed version of the newly-developed semi-analytical ＇finite element method of lines＇ （FEMOL） is proposed in this paper for accurate and efficient computation of stress intensity factors （SIFs） of two-dimensional notches/cracks. The circular regions surrounding notch/crack tips are taken as the complementary energy region in which a number of leading terms of singular solutions for stresses are used, with the sought SIFs being among the unknown coefficients. The rest of the arbitrary domain is taken as the potential energy region in which FEMOL is applied to obtain approximate displacements. A mixed system of ordinary differential equations （ODEs） and algebraic equations is derived via the sub-region generalized variational principle. A singularity removal technique that eliminates the stress parameters from the mixed equation system eventually yields a standard FEMOL ODE system, the solution of which is no longer singular and is simply and efficiently obtained using a standard general-purpose ODE solver. A number of numerical examples, including bi-material notches/cracks in anti-plane and plane elasticity, are given to show the generally excellent performance of the proposed method.

NUMERICAL SIMULATIONS OF CAVITATING FLOWS

A new model, which involves viscous and multi-phase effects, was given to study cavitating flows. A local compressible model was established by introducing a density-pressure function to account for the two-phase flow of water/vapor and the transition from one phase to the other. An algorithm for calculating variable-density N-S equations of cavitating flow problem was put forward. The present method yields reasonable results for both steady and unsteady cavitating flows in 2D and 3D cases. The numerical results of unsteady character of cavitating flows around hydrofoils coincide well with experimental data. It indicates the feasibility to apply this method to a variety of cavitating flows of practical problems.

URANS simulations of ship motion responses in long-crest irregular waves

In this paper, numerical prediction of ship motion responses in long-crest irregular waves by the URANS-VOF method is presented. A white noise spectrum is applied to generate the incoming waves to evaluate the motion responses. The procedure can replace a decade of simulations in regular wave with one single run to obtain a complete curve of linear motion response, considerably reducing computation time. A correction procedure is employed to adjust the wave generation signal based on the wave spectrum and achieves fairly better results in the wave tank. Three ship models with five wave conditions are introduced to validate the method. The computations in this paper are completed by using the solver naoe-FOAM-SJTU, a solver developed for ship and ocean engineering based on the open source code OpenFOAM. The computational motion responses by the irregular wave procedure are compared with the results by regular wave, experiments and strip theory. Transfer functions by irregular wave closely agree with the data obtained in the regular waves, showing negligible difference. The comparison between computational results and experiments also show good agreements. The results better predicted by CFD method than strip theories indicate that this method can compensate for the inaccuracy of the strip theories. The results confirm that the irregular wave procedure is a promising method for the accurate prediction of motion responses with less accuracy loss and higher efficiency compared with the regular wave procedure.