.png){kind=link}
Updated Lagrangian Explicit Finite Element Method (FEM) CPU/GPU based solver.
WeldFormFEM is aimed to solve solid mechanics large strain problems, such as metal forming.
The idea is to work via 2 different solvers:
1 - Pure lagrangian with rezoning (adaptive mesh refinement or AMR), Work In Progress as first option
2 - Coupled Eulerian (fixed mesh or Arbitrarian Eulerian Lagrangian) Lagrangian solver based on Benson works.
WeldFormFEM works both on Ubuntu and Windows.
you can select to build it to CPU and GPU only by changing a single CMAKE var.
- Structure Of Arrays (SOA) data arrangement which allows fast CUDA accesing
- Explicit time integration
- C++/CUDA CPU/GPU(WIP) Architectures
- Constant Stress Tetra/Triangle Element with Average Nodal Pressure (ANP) for volumetric locking fixing
- Reduced Integration Hexaheadra with viscous hourglass control
- OpenMP (WIP) CPU parallelization
- Contact algorithm (WIP)
- Automatic Remeshing (rezoning) (WIP)
- 2D Plain Strain/Axisymm & 3D
- Thermal-Mechanical coupling (WIP)
Locking (left) and fixing (right) tetra
Compression cylinder
Compression cylinder HEXA w/reduced integration - TETRA
Bending
Hexa/Quad Hourglass
CUDACXX=/usr/local/cuda-12.3/bin/nvcc cmake ../WeldFormFEM -DBUILD_GPU=ON
To update libraries (LSDynaReader and Math)
git submodule update --init --recursive
Link here #https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/ to see different architectures.
Cuurrently working axisymmetric with hourglass for area weight in F90 version
reduced_int = .True. call AddBoxLength(0, V, Lx, Ly, 1.0d0, r, rho, h,reduced_int)
!! AFTER ADD BOXLEN axisymm_vol_weight = .false. bind_dom_type = 3 !!!AXISYMM, AFTER CREATING BOX!
elem%sigma_y(:,:) = 300.0e6
do i=1,node_count print *, "NODE ELEMENTS " print *,"i count ", i , nod%elxnod(i),nod%nodel(i,:) end do
nod%is_fix(:,3) = .true.
Then in calc elem forces:
''' if Area weight #-------------
fa = 0.25d0/elem%radius(e,gp) * elem%detJ(e,gp) !!! THEN IS WEIGHTED BY 4 in case of gauss point =1
!!! AREA WEIGHTED, BENSON EQN 2.4.3.2
!!! 2.4.3.2 remains sig * Area/(4 r0), which is (4detJ)/(4r0) = detJ /r0
!!! LATER IS MULTIPLIED BY WEIGHT WICH GIVES THE AREA
elem%f_int(e,n,1) = elem%f_int(e,n,1) + elem%dHxy_detJ(e,gp,2,n) * elem%sigma (e,gp, 1,2) - &
(elem%sigma (e,gp, 1,1) - elem%sigma (e,gp, 3,3) ) * fa
elem%f_int(e,n,2) = elem%f_int(e,n,2) + elem%dHxy_detJ(e,gp,1,n) * elem%sigma (e,gp, 1,2) - &
elem%sigma (e,gp, 1,2) * fa
! print *, "fa ", elem%dHxy_detJ(e,gp,1,n) * elem%sigma (e,gp, 1,2)
! print *, "term 2 ", elem%sigma (e,gp, 1,2) * fa
'''