ImpressionFEA 0.9.0: installer, EULA, mesh check, ASM flash design

Downloads: …-full-setup.exe / …-light-setup.exe install to Program Files with a Start Menu entry and optional SOLIDWORKS add-in registration. The .zip packages are portable. Full includes the NVIDIA GPU solver; light is CPU only (the GPU solver can be added later). Each file has a .sha256 checksum.

License: free for personal and educational use; commercial use requires a Commercial License. ImpressionFEA is an educational and conceptual tool, not intended for professional engineering use. See EULA.txt.

• End User License Agreement (EULA.txt): free for personal and educational use; commercial use requires a
  Commercial License (30-day evaluation allowed). ImpressionFEA is an educational and conceptual tool, not intended
  for professional engineering use, because its modeling assumptions are substantial. The EULA also covers warranty and
  liability, privacy (nothing is sent anywhere) and the bundled NVIDIA and Intel libraries. The installer asks you to
  accept it; the source code stays under CC BY-NC 4.0. Reports now state that results are preliminary estimates from a
  simplified model, not for professional engineering use. Results published under the free licence credit ImpressionFEA;
  commercial results shared outside the licensee keep the report's ImpressionFEA identification.

• Windows installer (…-win-x64-full-setup.exe / …-light-setup.exe, next to the zip packages): installs to
  Program Files with a Start Menu entry (desktop shortcut optional) and an uninstaller. The optional SOLIDWORKS
  add-in (ticked when SOLIDWORKS is found) is registered by the installer itself, machine-wide, without PowerShell;
  uninstalling removes it. The light installer can download the NVIDIA GPU solver on its last page.

• Automatic element size: in 3D, h is chosen for about 8,000 nodes in the initial mesh (Advanced ▸ Target
  nodes, remembered between sessions), calibrated on the actual mesh (the best of up to three trial meshes, since a
  coarse mesh counts nodes in lumps), within the node budget and the four-elements-
  across limit. It follows every change of the workpiece, so a resized die-only workpiece no longer disappears for
  having too few or too many elements. Automatic values are grey; typing a value makes it a bold override that
  stays; clearing the box returns to automatic. The stroke increment follows h the same way (0.2·h). Setups record
  whether h and Δs were automatic; older setups and the demos keep their stored values as overrides.

• Input tooltips show every field's allowed range in the display units; the element-size tooltip also gives the
  valid h range for the current workpiece and the h for the target node count. The mesh read-out shows the actual
  node and tetrahedron count of the preview mesh.

• Errors and warnings moved from the bottom of the configuration panel to a box at the top left of the view:
  errors (red: the run cannot start and the workpiece is not drawn, with the reason) and warnings (amber), click to
  fold. A new message lists both limits when no element size fits the workpiece.

• ETA in the execution bar while solving, weighted towards the recent increments (they slow down near die
  closure); the tooltip gives the finish time and, while a queue runs, the time for the remaining items.

• 3D contact accuracy: contact tolerances now scale with the local element size instead of the global h, so
  refined elements no longer sit inside the dies. On the hex bolt head the deepest node penetration at closure fell
  from 0.45 mm (41 nodes) to under 0.2 mm, and runs got slightly faster. Only the last recovery passes of an increment
  that would otherwise fail fall back to the old, looser tolerances.

• 3D solver about 2.4× faster: the element work of each Newton iteration runs on all CPU cores (results summed in a
  fixed order, so they do not depend on the core count), and the preconditioned iterative solve backs off after
  repeated failures instead of costing nearly a factorisation each time. The Balanced hex bolt head went from about
  6.4 to 2.7 minutes. Every 3D log ends with a time profile (assembly, factorisation, iterative solve, contact,
  remeshing).

• Stroke increment follows the flash: each increment compresses the thinnest layer held between the dies by at
  most 10 %, so the step shrinks ahead of the flash instead of failing and being cut back.

• Hex bolt head preset is a quarter model (it is symmetric about x and y) with a 1 mm radius on the die edge into
  the flash land. At the same run time the mesh is much finer: the workpiece volume is now conserved within about 1 %
  (the full model lost 6–7 % into the thin flash, which understated the load by about a third to a half). Demos 4 and
  8 use the automatic element size.

• Mesh quality (Advanced ▸ Mesh quality 3D): Fast (about 4,000 nodes), Balanced (8,000, default) or Accurate
  (20,000, node budget 40,000) for your own parts when run time matters less; remembered between sessions. A 3D run
  that changes the workpiece volume by more than 3 % ends with a coarse-mesh warning.

• Volume conserved on remeshing: the free surface (in the end the flash rim) may now move up to one element per
  remesh to restore the lost volume, instead of a quarter: the coarse hex bolt head now ends within about 1 % instead of
  −6 %. Volume was not what made coarse meshes low on load, though:

• Flash-resolution check: when the dies close with less than about one element through the thinnest flash, the run
  ends with an accuracy note (the flash pressure, and so the load, is likely understated) and a suggestion to use the
  Accurate mesh quality or a smaller element size.

• 3D Newton solver about 30 % faster: a blended tangent (steadier for low strain-rate sensitivity), a step length
  from a quadratic model instead of repeated halving, and switching to a secant correction when Newton keeps
  overshooting; convergence tolerance 1e-3 (was 1e-4). Loads stay within the solver's own run-to-run scatter. Older
  setups keep the tolerance they were saved with.

• Reproducible results: the choice between an iterative solve and a refactorisation no longer depends on timings,
  so on the CPU solver (MKL) the same setup gives the same answer however busy the computer is (it varied by 1–2 %
  before). GPU (cuDSS) factorisations are not bit-reproducible. Identical GPU runs of the hex bolt head differ by up
  to about 1 %, and with the former 1.6 mm flash GPU and CPU differed by 4 %. Auto therefore keeps the CPU unless
  the GPU is at least 1.5× faster on the first factorisation (it took the faster of the two before, which could change
  from one run to the next). Choose the CPU solver under Advanced when reruns must match exactly on any computer.

• Closing load: run summaries, the completion message and comparison reports show the load at die closure from a
  straight-line fit over the last millimetre of stroke. It scatters by about ±1 % between equivalent runs, against
  ±3 % for the single closure increment, so it is the figure to compare designs and meshes by.

• Mesh check… (RUNS, 3D): is the selected run's mesh fine enough? Every 3D run stores its state at a 2 mm die gap.
  The check re-solves the rest of the stroke at the run's element size and at 0.8× and 0.64× of it, in parallel. It
  estimates the closing load at a perfectly fine mesh (settled finer levels, or a Richardson extrapolation). It says
  how far the run is from it: converged within 3 %, not converged, still changing, or inconclusive. It shows progress
  and an ETA, and Abort stops it. The result goes in the status line, the log, the run summary and the reports. It
  takes roughly 3–10 times as long as the run. On the Balanced hex bolt head (2 mm flash) it finds the run about 9 % or
  more low, in 7 minutes. With the former 0.8 mm flash the run was 40 % or more low and the check took 30 minutes.

• Flash lands redesigned to the ASM guidance (thick side): every built-in die's flash now follows Altan, Ngaile &
  Shen (§14.4–14.5). Flash thickness is 1.25–2× the industry-survey trend for the part's steel weight (Fig. 14.13),
  with a land ratio w/t of about 3 (the load data of Fig. 14.14 cover 2–4). The thick side keeps the isothermal results
  steady and the meshes affordable; thermal effects, when added, will change flash design and loads substantially.

  • Gear blank: 3 mm flash, land shortened to 9 mm (w/t 3.0; was 3.5). Load and fill are unchanged: the flash reaches
    the gutter at closure either way.
  • Flanged shaft: 2 mm flash, 6 mm land (was 7.2).
  • Connecting rods (3D, and the 2D shank section): 2.5 mm flash with a 7.5 mm land, the book's own connecting-rod
    flash (was 1.8 and 1.6 mm, w/t 4.4 and 4.0). The 3D rod's straight bar now carries 30 % excess stock (was 10 %):
    it fills about 91 % at closure (was 86 %); a straight bar cannot fill a rod completely without a preform.
  • Hex bolt head: 2 mm flash (was 1.6) with a 1/4" land (5.8 mm flat, w/t 2.9; was 3/16"). It still fills
    (99.8 %). The Balanced closing load is 95.6 t on the CPU solver (was 105.6 t). With the thicker flash, GPU and CPU
    runs agree within about 0.6 % (was 4 %).
  • Demo 8 now sweeps 1.6, 2.0 and 2.6 mm instead of 0.8, 1.6 and 3.2 mm, which lay outside the guidance.

  FlashDesign (Core) holds the digitised survey curves, Vieregge's equations (14.6, 14.7) and a guideline check.
  Each die now declares its flash land, and the tests measure it on the die geometry. The README has a new section,
  "Flash design of the built-in dies".

• Staged resolution and graded meshes (solver options, for studies): a "fine finish" that switches to a finer
  element size near die closure, and graded isosurface stuffing (fine where the metal deforms, coarse in rigid
  zones). Both work and are tested; neither is in the user interface yet.

• Handbook load estimate (Core): the simplified slab method of Altan, Ngaile & Shen (Eqs 14.8–14.11, with the
  sliding and disk-shearing cavity modes of Table 15.1), with benchmark tests against the book's worked examples.

• Crash log moved to %LOCALAPPDATA%\ImpressionFEA; your own and edited demos go in Documents\ImpressionFEA\demos
  (a file there replaces the built-in demo of the same name) — the install folder can be read-only.