This is a notebook of laser-cutting experiments.
So far, I’ve only achieved some super basic stuff.
The purpose here is rather like the objective of RepRap: design a machine that can make itself. RepRap is a very inspiring project which moves us a big step forward toward the glorious post-scarcity future of universally accessible material abundance through self-reproducing machinery and freely shared information.
However, I think RepRap’s focus on additive fabrication with FDM may be a big handicap: while FDM, like other additive processes, lets you do fairly free-form fabrication, the materials available are relatively limited in their properties, and in particular, they are inherently incapable of fabricating certain parts of the printer itself: the hotend especially, but also kind of the heated bed. And FDM is very slow, both in terms of absolute quantity of material and in terms of information bandwidth. And its tolerances are not that great, especially in the Z dimension.
Other additive processes, like powder-bed printing (whether binder-based or sintering-based), Carbon3D and other optical polymerization systems (including the original SLA), and especially processes using DLP, have better bandwidths.
There’s another family of digital fabrication processes which are more limited in the shapes they can achieve than the additive processes mentioned above, but which I think may actually work better for RepRap’s aim of a personal-scale, environmentally sustainable self-reproducing factory. These processes work by cutting prefabricated sheets of material into pieces with a precise kerf at right angles to the surface following a two-dimensional design.
I can’t find a general term for this family of processes, so I’m going to call them “planar fabrication processes”, even though that already means a different process in semiconductor fabrication. The most accessible of these processes is laser cutting, but to some extent these processes are interchangeable: they have very similar capabilities.
Planar digital fabrication processes share FDM’s ability to produce custom shapes from digital files at high bandwidth, but they have the great advantage that they only need to apply energy to the edge of the shape they’re creating, rather than its entire volume. They have a lot of scope for optimization processes and parametric design, but work at one or two orders of magnitude higher speed, and they work with a much larger range of materials, including materials with two to three orders of magnitude better properties than the PLA and ABS used in FDM.
The only software I’ve found so far for planar digital fabrication in general is SheetCAM.
(Maybe “panel cutting” or “sheet cutting” is a better name for this category? “Panel cutting” seems to be an established term actually.)
The most accessible of these processes is laser cutting of plastics and wood-based materials; although an entry-level Epilog laser cutter costs US$8000, service bureaus throughout the urban world have these machines available and are willing to do quantity-1 runs for costs under US$10.
The particular laser cutter I’ve tried so far will cut MDF sheets of up to 810×450 mm with about 60μm precision and a kerf of about 100μm, it has no minimum corner radius beyond that implied by the imprecision of the kerf itself, and the kerf can start or stop anywhere. The laser cutter can additionally mark surfaces or engrave (“gouge”) them to different depths.
MDF’s material properties are actually fairly similar to PLA, except that MDF has about a third the tensile strength, doesn’t melt, and falls apart in water. But you can shape 100 grams of MDF on a laser printer in about the time it takes to shape 1 gram of PLA on a Prusa Mendel, depending on how elaborate your shape is. And the same laser cutter can also cut PMMA, which is several times stronger and transparent to boot.
Several other processes share many of the same properties.
Higher-powered laser systems can cut metals too; while a laser cutter for cutting MDF and plastics might have a laser of 30 to 80 watts output, a laser cutter for metals might have an output of 250 to 4000 watts, depending on the thickness, which means that the power to run the laser is kilowatts up to tens of kilowatts. I don’t know much about these machines, and I’ve never seen one.
The lower-power machines I mentioned above can, at best, cut metal foil.
I don’t have any experience with these yet, although I’ve used a manual plasma cutting torch, which costs only about US$1000.
A CNC plasma cutter moves a plasma cutting torch over a cutting table to cut planar shapes in a sheet of metal. Instead of cutting the shape with light, it generates a plasma arc between an internal tungsten electrode and the workpiece, which is blown through a water-cooled nozzle and the workpiece with a high-pressure air supply. This is a lot more efficient than a laser, since nearly all of the energy ends up being applied to the workpiece, but it’s also a lot less precise, producing a much rougher cut surface that’s not as perpendicular.
An entry-level plasma cutting table seems to cost about twice as much as an entry-level laser cutter, at about US$15000 or US$20000. The Global Village Construction Set is working on an open-source torch table, and Jacques Mattheij has open-sourced his own torch table design at http://jacquesmattheij.com/how-to-build-a-windmill-ii, which he used to build his own windmill power plant.
Plasma cutters are also capable of gouging surfaces and making cuts at non-perpendicular angles to the workpiece surface.
They have some disadvantages. They can’t cut nonmetals because the arc needs to terminate on the workpiece. The moving nozzle has more mass than the mirrors of a laser cutter, and the plasma arc has to continue running at a somewhat constant power in order to maintain a consistent kerf, and consequently at least some of these machines have a minimum turning radius. Sharp outside corners are achievable even then at the cost of a sort of cloverleaf approach. I think that when you start a new kerf, it starts out a bit wider and sloppy. Finally, they generate a heat-affected zone near the kerf, altering the material properties.
Metals have a lot of advantages over plastics and MDF for many applications: they are on the order of 100 times stronger and stiffer, have about 10 times the strength-to-weight ratio, and have about a tenth of the thermal coefficient of expansion. Their main disadvantages are that they are more expensive and harder to shape.
I have no experience with this either.
Wire EDM also cuts a metal workpiece by vaporizing it with plasma, but instead of blowing the plasma through the metal with air pressure, it generates the plasma in many tiny sparks underwater in between a tiny brass wire and the workpiece. The water cools the workpiece and minimizes the heat-affected zone; the wire is continuously fed through the workpiece in order to get rid of the wire that has already been exposed to plasma.
Wire EDM is a much more precise process than those mentioned above, with a typical kerf width of some 300μm and precision of better than 10μm. The process can cut arbitrarily tight corners, but can’t start a cut in the middle of the workpiece unless you drill a starter hole first and then feed the wire through it.
Homebuilt hobbyist wire EDM machines are all over YouTube, but I don’t think they get very good precision. Commercial wire EDM machines seem to be relatively inexpensive, starting around US$1000. They have higher operating costs than the other machines mentioned above, though, because of the brass consumption. As a result, I don’t think wire EDM is a great bet for self-reproducing machinery.
Scrollsaws cut curved kerfs in wood workpieces using a thin reciprocating saw blade. They are far and away cheaper than any of the above-mentioned processes, starting at around US$100, but also less precise, and with a kerf width of as much as 1mm. As with wire EDM, you have to predrill holes if you want to start a cut in the middle of the workpiece. The blade width imposes a minimum curve radius, but the biggest issue is that typically you have to rotate either the workpiece or the saw as you curve, and consequently CNC scrollsaws are apparently unknown. Spiral scrollsaw blades, which helically cut in any direction, are available; I think they cut a wider kerf than a regular scrollsaw.
However, if you have a design for planar fabrication, a totally ghetto alternative to laser cutting is to print it out on a printer, glue or toner-transfer the printout onto wood, and cut it out by hand with a scrollsaw, jigsaw, or coping saw.
A CNC router is essentially a CNC milling machine. Rather than reciprocating up and down like a scrollsaw blade, the router bit rotates around a vertical axis, and so it can cut equally well in any horizontal direction. CNC routers for cutting wood are available from about US$500, and they have better precision than scrollsaws, but they leave much wider kerfs. Because the bit cuts a round hole, inside corners have a minimum radius of half the kerf width.
However, CNC routers are capable of extensive non-planar fabrication as well, particularly if you vary the bit shape.
These are very noisy machines.
I have no experience with these. They’re fairly new, dating only from the 1980s, and still extremely expensive.
A waterjet cutter sends a supersonic jet of water a few tens of microns wide, typically formed by a tiny orifice in a synthetic ruby, through your workpiece, dragging fine abrasive with it. Waterjets can cut any material: paper, wood, MDF, steel, titanium, granite, Invar, beryllium, glass, quartz, tungsten carbide, you name it. When this process isn’t completely submerged, it often throws off sparks, despite the water. The only requirement is that the abrasive be at least as hard as the material. Garnet, the standard abrasive, is about as hard as quartz; cutting ultrahard materials like silicon nitride or ruby requires using a harder abrasive such as ruby (aluminum oxide). I think cubic boron nitride, boron carbide, and diamond can be cut using abrasive waterjets, but they would require one of those three materials themselves as an abrasive — and I can’t find references to anyone doing it.
Waterjet cutters often are capable of varying the cutting angle as well (“multi-axis cutting”) but, like the other processes described above, can’t limit the depth of cut.
Waterjet cutting doesn’t produce a heat-affected zone, doesn’t have a minimum turning radius, typically has a kerf of around 1mm with a precision of around 100μm, more precise than any of the other processes mentioned above except for wire EDM. The kerf, as with most of these processes, is tapered; in this case the far side of the cut is narrower.
These machines are more expensive than the others mentioned above, with even entry-level machines starting above US$100k. Worse, their operating expenses are also high, as they consume massive quantities of garnet abrasive and energy. Building a waterjet cutter requires high-pressure plumbing and exotic materials. For these reasons, I don’t think waterjet is currently a viable way to build a small-scale self-reproducing factory.
An ordinary design is a single, fixed shape and set of dimensions. A parametric design is a potentially infinite set of shapes of different dimensions; it can be adapted to different circumstances. For example, BoxMaker is a parametric design for a right rectangular parallelepiped to be cut on a laser cutter, presented as a web site.
In some cases, a parametric model is defined with certain parameters as inputs — in the case of BoxMaker, the parameters are width, depth, height, units, material thickness, notch length, cut width, and a couple of booleans — while others are outputs. Such a parametric model can be thought of as merely a procedure which computes a design given those inputs.
OpenSCAD is a popular system for such procedural parametric design, commonly used for FDM design (most of the parametric designs on Thingiverse are designed with OpenSCAD), but it is very limited; because the model procedures have no access to the generated geometry, it’s impossible to do something as simple as a fillet or a gusset in a generic way, and often quite difficult to do it for a given design.
Some more advanced parametric modeling systems, like that in SolveSpace, use a more flexible “constraint satisfaction” paradigm. A constraint-satisfaction parametric model consists mostly of a set of constraints relating its various parameters. For example, a square might have eleven parameters: upper left, upper center, upper right, left center, center, right center, lower left, lower center, lower right, side length, and rotation. From any three of these parameters, and any two other than side length and rotation, the other nine can be computed. Although parametric constraint models are more flexible than procedural parametric models, the adoption of such systems is somewhat impeded by the difficulty in making them work and figuring out why they don’t work when they don’t.
Designing parametrically is generally a recursive process: a parametric design includes smaller parametric designs within it. For example, a planar spring snap-fit joint has a design parameterized by the loads it must bear and perhaps safety factors and material properties; but a parametric model of a three-dimensional box that snap-fits together might contain a number of these snap-fit joints within it.
The enormous potential of digital fabrication is to enable design through optimization, a word I use in a fairly technical sense here. The “optimization” problem is, given
- a multidimensional design space;
- constraints on feasible designs; and
- an objective function to minimize;
to find the point in the feasible subset of the design space that minimizes that objective function. For example, to build a bridge, your design space might be all of the different possible ways the available materials could be connected together, you are subject to constraints of connecting to both sides of some space, supporting a given load and resisting wind and so on, and you might want to minimize the cost of building the bridge subject to those constraints.
An already-mainstream example of this approach is “topology optimization”, in which typically the dimensions of the design space are density of a single material in different cuboidal finite-element-analysis cells of a physical space, the constraint is typically a maximum amount of total material to use, and the objective function is typically stiffness (or rather its reciprocal, compliance.) There are a variety of existing proprietary software packages that do this; solidThinking Inspire, Abaqus, ANSYS, HyperMesh, and Midas NFX are some examples. But topology optimization research continues, exploring constraints and objective functions touching on manufacturability, thermodynamics, and computational fluid dynamics.
Except in special cases where the problem has low dimensionality and strong continuity properties, finding the actual global optimum is computationally infeasible. If I understand correctly, topology optimization is typically done with simple gradient descent. Other general optimization algorithms that could be applied include simple random sampling, gradient descent with random restart, genetic algorithms, tree search, pathfinding, and non-chronological backtracking, along with more advanced forms of design-space sampling that attempt to generate “interesting” starting points by recombining patterns that are already known to work under some circumstances. Several of these approaches can get orders-of-magnitude speedup from incremental computation.
(Topology-optimization results often look nothing at all like human designs, resembling trees, corals, jellyfish, or bones more than traditional industrial designs.)
Another already-mainstream example is “nesting software”, which optimizes the use of a piece of stock by choosing where to cut things out from it so that they fit together well with a minimum of waste. This is commonly used in panel cutting, for example.
But all of these are iterative algorithms: they take a proposed design, evaluate it in simulation, and use the results to come up with more proposed designs. Doing this for the design of a physical thing requires a sufficiently accurate simulation of the performance of the thing to make the optimization process useful rather than counterproductive.
The more of the design and fabrication process that the optimization process can model, the better it can do. For example, if it knows the limitations and imprecisions of the fabrication process, it can work around them; to the extent that it can model the production cost, it can reduce it rather than increasing it. Current mainstream topology optimization systems are still far too primitive to take into account such things. But these optimization concerns cut across these traditional abstraction layers; to the point that you can allow the optimizer to cut across those layers, it can produce a much better design.
As a very simple example, if you have a set of parts to laser-cut from a sheet of MDF, an arrangement on the sheet that nestles them together as much as possible will use less MDF and require much less cutting. If the optimization process is also capable of altering the shapes of the parts, it can make them fit together a great deal better — but that only works if it knows what objective the part shapes were intended to achieve.
Optimization is in some sense a generalization of constraint satisfaction: constraints are absolute, all-or-none requirements, while optimization seeks to fulfill requirements as much as possible, including greatly exceeding minimal requirements, if possible. You can define an objective function to take into account as many constraints as needed.
Right now I’ve successfully done some really basic tests and very simple constraint-propagation-driven gear designs.