DiffusionNet: Accelerating the solution of Time-Dependent partial differential equations using deep learning

  • 2020-11-19 18:28:42
  • Mahmoud Asem
  • 10


We present our deep learning framework to solve and accelerate theTime-Dependent partial differential equation's solution of one and two spatialdimensions. We demonstrate DiffusionNet solver by solving the 2D transient heatconduction problem with Dirichlet boundary conditions. The model is trained onsolution data calculated using the Alternating direction implicit method. Weshow the model's ability to predict the solution from any combination of sevenvariables: the starting time step of the solution, initial condition, fourboundary conditions, and a combined variable of the time step size, diffusivityconstant, and grid step size. To improve speed, we exploit our model capabilityto predict the solution of the Time-dependent PDE after multiple time steps atonce to improve the speed of solution by dividing the solution intoparallelizable chunks. We try to build a flexible architecture capable ofsolving a wide range of partial differential equations with minimal changes. Wedemonstrate our model flexibility by applying our model with the same networkarchitecture used to solve the transient heat conduction to solve the InviscidBurgers equation and Steady-state heat conduction, then compare our modelperformance against related studies. We show that our model reduces the errorof the solution for the investigated problems.


Quick Read (beta)

loading the full paper ...