GNAL/docs/README.md

16 KiB

Development Notes

  1. Dependencies
  2. Build Scripts
  3. Version Notes - V1
  4. Version Notes - V2
  5. Version Notes - V3
  6. Benchmarks

This project can be edited with only OpenSCAD and the source files in the scad/*_v1, scad/*_v2 or scad/*_v3 directories which make reference to files from scad/libraries. If you wish to run the development scripts you should install the following dependencies.

With just OpenSCAD, you can use scripts such as scad/50ft_v3/gnal_50ft.scad and export the different modules in OFF, AMF, 3MF, DXF or SVG (drawing) format. The CSG models can be imported into FreeCAD and DXF models should be readable by AutoCAD and QCAD.


Dependencies

Build Scripts

Running any of the build scripts scripts--scripts/v1.sh, scripts/v2.sh or scripts/v3.sh--will start an OpenSCAD build process of all components and will log stats about the resulting files and render times to notes/v1.csv, notes/v2.csv or notes/v3.csv.

Keep in mind that V1 and V2 compile times are extremely long and all scripts will use an entire CPU core at 100% utilization while rendering. It's best to run these scripts in the background on a powerful machine or better yet, not at all. See the stl folder for pre-compiled STL files for 3D printing or the releases page for .zip and .tar.gz archives of all versions.

These scripts will render STL files, PNG images of the files and capture metadata about the render process while doing so.


Version Notes

V1

Intended to be mostly compatible with existing processing spirals with some caveats. A spacer that is typically threaded has been replaced by a friction fit part so they are not interchangeable. This version is designed to fit in existing tanks and use the same spindle screws.

In the process of building this first version several approaches were evaluated to generate the spiral shape. The first is what's best described as a brute force approach laying out an excessive amount of rectangular facets and unioning them together in such a way that the result would be a single continuous spiral. This took hours to days to render depending on the machine used.

Besides the exhausting render times this approach bugged me for one reason: all facets of the spiral were the same size, meaning that the small diameter inner parts of the spiral were packing in millions of unnecessary polygons to allow for the large diameter parts of the spiral to be smooth. This didn't sit well with me. How many CPU hours are being burned by adding detail to a place that doesn't matter? Answer: a lot.

Finally, an external library called path_extrude.scad by @gringer was brought in to handle the complicated spiral extrusion step. A simple function that plots a spiral in Cartesian coordinates is used to draw the path and a 2D triangle is extruded along it by the library. This allowed for the path to be drawn at a consistent "resolution" throughout the entire spiral, so the facets of the outermost and innermost parts were the same or extremely similar.

Here is that function reduced to a single line in order to generate an array of coordinates.

spiralPath = [ for(t = [0 : $fn + 1]) [((d / 2) + (t * increment)) * cos(t * angle_i), ((d / 2) + (t * increment)) * sin(t * angle_i), 0] ];

The experimentation in this version predate this particular git repo and so will not be found in the git history, but you can find the vestigial functions in the spiral directory used for benchmarking different approaches.

Beware

This version of the spiral must be printed with supports. The spirals themselves are suspended over voids and this particular feature is addressed in the later versions.

Rendering

Rendered using OpenSCAD version 2019.05 on a 2.2 GHz Core i7 (I7-4770HQ) chip running macOS 10.14.

Model Size (bytes) Facets Volume (mm3) Render Time (sec)
gnal_50ft_spacer.stl 991452 5736 2888.155029 68
gnal_50ft_top.stl 2132181 12624 57936.746094 233
gnal_50ft_spiral_top.stl 36509561 214404 120299.773438 12249
gnal_50ft_spiral_bottom.stl 36606204 214970 121519.937500 13698
gnal_100ft_spacer.stl 991452 5736 2888.149658 74
gnal_100ft_top.stl 3302563 19552 102590.546875 477
gnal_100ft_spiral_top.stl 92423369 542836 223602.078125 89137
gnal_100ft_spiral_bottom.stl N/A N/A N/A N/A

V2

This version aims to improve printability over the V1 model and reducing render time of the spiral. The biggest change to the physical structure of the design is the removal of overhangs from beneath the spiral film guide.

When printing a model with FFF printing, any piece that overhangs empty space (usually) needs to be supported by a temporary removable structure beneath it otherwise you risk the piece drooping. In the case of V1 model, the spiral was completely suspended by the spokes of the reel with large gaps of empty space. This means there were 90 degree overhangs under the most critical part of this model; the grooves for holding the film in position. Printing and removing support structures from beneath the fragile spiral made post-production dangerous for the piece and time consuming.

The solution to this was to extend the spiral to the bottom of the reel and remove triangular sections from them to allow for a lighter print and better chemistry movement. Most printers should be able to print these structures without any support material or any resulting deformations in the model.

One other change in this version is that it reduces the spiral models to a single one to be duplicated, rather than two distinct top and a bottom pieces that differ in only minor ways. This decision was motivated by an interest in making this design better (cheaper) for injection molding.

A secondary benefit of reducing the spiral to a single model was to immediately cut render times for the entire project nearly in half before any other optimizations were made. The first meaningful code optimization toward this goal was provided by a helpful comment made on a long-forgotten design shared on a 3D printing forum.

@sousvide59 (Les Smith) writes

It may be more efficient to approximate the spiral as a series of arc segments, like this <Github gist>.

Les was right. This reduced the several hours render time to 1-2 hours, which worked for this version. Ideally this will be improved further in future versions. Beyond some explorations into OpenSCAD hacks (rendering each complete rotation of the spiral and stitching all resulting STLs) the next version will incorporate other languages and platforms to find the fastest render time for a GNAL spiral. All previous approaches are being compiled into a suite of tests to benchmark render times.

Rendering

Rendered using OpenSCAD version 2020.01.17 on a 3.2 GHz Core i5 (I5-4460) chip running Ubuntu 18.04.

Model Size (bytes) Facets Volume (mm3) Render Time (sec)
gnal_50ft_spacer.stl 991452 5736 2888.150879 22
gnal_50ft_top.stl 2132181 12624 57937.210938 73
gnal_50ft_spiral.stl 34628449 193450 178181.250000 2341
gnal_50ft_insert_s8.stl 5228272 27230 3493.560303 97
gnal_50ft_insert_16.stl 7922994 41426 4664.952637 155
gnal_50ft_spacer_16.stl 561267 3272 4015.912109 19
gnal_100ft_spacer.stl 991452 5736 2888.152100 23
gnal_100ft_top.stl 3302563 19552 102590.812500 118
gnal_100ft_spiral.stl 59279238 330000 345431.531250 4542
gnal_100ft_insert_s8.stl 5228272 27230 3493.559326 99
gnal_100ft_insert_16.stl 7922994 41426 4664.937500 160
gnal_100ft_spacer_16.stl 535264 3112 3964.118164 17

V3

The goals of V3 are to greatly optimize the spiral generation code for speed and to restore the feature of the V1 spiral which maintains a consistent size of individual facets throughout the spiral even as the diameter changes. This will be considered a stable release candidate for publishing the project.

Since the benchmarking process (see below) was developed between V2 and V3, render times are optimized in this iteration of the project. The success of the spiral_3.scad approach stood out from the rest as fastest, so it was reworked into what exists in V3.

The spiral itself is plotted in 2D with a relatively simple formula that is expressed in this OpenSCAD script through a number of in-line helper functions. It draws the position of various points along the spiral path and then uses the path_extrude.scad library to extrude a shape along those coordinates. This proves to be fast and efficient while not sacrificing any detail in the geometry.

Prior to release a serious flaw was found while printing the V3 design. The attempt to remove the need for supports in V2, actually printing the spiral was creating curious side effects during fabrication. Not having material in the voids below the spiral, it seems, allowed air to cool the part and would consistently cause prints to fail when it reached the actual spiral at the top of the reel. Since this is the most important element of the reel, the triangles have been removed and the design is more similar to V1. Testing continues on this version.

In a compromise to make the process of removing the support material less dangerous to the detail on the top, the spiral itself extends lower than V1 into the space between the spokes of the reel. The spiral is also made thicker and is a multiple of my nozzle diameter (0.4mm). This uses slightly more material but is less fragile than V1 and test prints proved that the supports were less difficult to remove than in the earliest model.

This version will also contain a 4x reel stacking feature so that all models can be stacked with 4 spiral reels and a top piece. That will give 200ft capacity to the 50ft model and 400ft capacity to the 100ft model. A stretch goal for this version is to make a 35mm spacer and spindle set so that movie film in the format can be processed in 100ft lengths.

Beware

This version also requires the use of support material while printing.

Rendering

Rendered using OpenSCAD version 2020.08.18 on a 2.2 GHz Core i7 (I7-4770HQ) chip running macOS 10.14.

Model Size (bytes) Facets Volume (mm3) Render Time (sec)
gnal_50ft_spindle_bottom.stl 3760384 75206 4134.077637 1291
gnal_50ft_spindle_top.stl 6915384 138306 22229.814453 1128
gnal_50ft_spacer.stl 286884 5736 2888.150635 62
gnal_50ft_top.stl 1104884 22096 57933.800781 585
gnal_50ft_spiral.stl 9500384 190006 171712.140625 1111
gnal_50ft_insert_s8.stl 1361584 27230 3493.544922 276
gnal_50ft_insert_16.stl 2071384 41426 4665.019531 439
gnal_50ft_spacer_16.stl 602084 12040 4019.470703 281
gnal_100ft_spindle_bottom.stl 3760384 75206 4134.064941 1275
gnal_100ft_spindle_top.stl 6979184 139582 22229.773438 1139
gnal_100ft_spacer.stl 286884 5736 2888.143555 63
gnal_100ft_top.stl 1620084 32400 102557.437500 998
gnal_100ft_spiral.stl 18364384 367286 326573.812500 3746
gnal_100ft_insert_s8.stl 1361584 27230 3493.548340 272
gnal_100ft_insert_16.stl 2071384 41426 4664.790527 450
gnal_100ft_spacer_16.stl 755684 15112 4019.479248 368

Benchmarks

In the process of publishing this repository I started questioning claims I was making in this README. Throughout the development of this processing reel I've been plagued by long render times. As a sanity check, I went through my personal development history on this project and produced 6 distinct spiral generation scripts that I ran through a series of tests to benchmark the render performance, total volume generated and number of facets produced. Render time was the primary metric that concerned me, but I considered the others important in comparing these different approaches.

This work led to the creation of the approach in spiral_7.scad and was ultimately used in V3.

An example of a single test pulled from the notes/benchmark.csv results. These example results are rendered using OpenSCAD 2020.05.23 on a 2.3 GHz Xeon Gold 6140 chip running Ubuntu 18.04.

Spiral Test Diameter (mm) Rotations $fn Size (bytes) Facets Volume (mm3) Time (sec)
spiral_1.scad 47 10 100 7409653 41064 5391.819336 209
spiral_2.scad 47 10 100 15349620 86646 3639.441162 855
spiral_3.scad 47 10 100 1336635 8004 3589.485596 0
spiral_4.scad 47 10 100 1607691 9624 3830.134521 23
spiral_5.scad 47 10 100 4711486 28188 3602.101562 8
spiral_6.scad 47 10 100 4265376 25396 14337.455078 120
spiral_7.scad 47 10 100 990006 5924 3581.499756 0

As you can see, the different approaches lead to wildly different render times for the same test. If you look at the complete results you will see many tests did not even finish due to exhausting memory on the machine or the process being killed by the cloud host (using too much CPU for too long, most likely). In the case of spiral_3.scad and spiral_7.scad measuring "0" seconds, this just means that the process finished rendering in less than 1 second.

In the scad/spiral directory you will find each individual script in a scad/spiral/spiral_#.scad file. The scripts/benchmark.sh script will render spirals at various resolutions and rotation counts and record the results in notes/benchmark.csv.

Insights gleaned from these benchmarks was incorporated into the latest versions of design. Consider this comparison of just the 50ft spirals (top spiral from V1).

# Model Size (bytes) Facets Volume (mm3) Render Time (sec)
V1 gnal_50ft_spiral_top.stl 36509561 214404 120299.773438 12249
V2 gnal_50ft_spiral.stl 34628449 193450 178181.250000 2341
V3 gnal_50ft_spiral.stl 9500384 190006 171712.140625 1111

Render times have gone down dramatically between V1 and V2. Times halved again between V2 and V3. The volume has stayed consistent with major changes in geometry (between V1 and V2). The file size of the V3 spiral has reduced to about 30% of the V1 and V2 spirals and the facet count remains roughly the same throughout (which was a surprise).

Faster render times mean more iteration and less time between tests. The next part of the process to examine is slicing which has primarily been done with Cura, but other engines will be looked at for their speed, efficiency and print quality.