Custom Scan Strategies in SLM / L-PBF with PySLM: Sinusoidal Scanning

Building upon the previous post that provided a detailed breakdown for creating custom island scan strategies, this further post documents a method for deploying custom ‘hatch’ infills. This is particularly desirable capability sought by researchers and has been touched upon very little in the current research. The use of unit-cell infills or in particular fractal filling curves such as the Hilbert curve have been sought for better controlling the thermal history and melt pool stability of hatch infills.

This has been previously explored in SLS [1]][2] and in SLM on a previous collaborators at the University of Nottingham investigating Fractal scanning strategy [3][4].

Typically, hatch infills are sequences of linear lines that form the the ‘hatch’ pattern. Practically, these are very efficient mechanism for infilling a 2D area by using 1D line elements when rastering a laser. Clipping of lines within polygons is intuitive. As discussed there are various scan strategies that can be employed to generate variations on this infill – i.e. stripe, checkerboard/island scan strategy and also modifying the order or sorting of the hatch vectors.

Geometrical scan strategies that adapt the infill based on the underlying geometry, i.e. lattices are acknowledges as ways for drastically improving the performance and the quality of these characteristic structures. This would be based on some medial-axis approach. This post will not specifically delve into this, rather, demonstrate an approach for custom infills on bulk regions.

Ultimately, drastically changing the behavior of the underlying hatch infill has not really been explored. This post will demonstrate an example that could be employed and explored as part of future research.

Custom Sinusoidal Approach

Sinusoidal scanning has been employed in welding research [5] and also in direct energy deposition (DED) [6][7][8] in order to improve the stability and quality of the joining or manufacturing process.

The process of generating this particular scan strategy requires some careful thought to improve the efficiency of the generation, especially given the overall increase in number of points require to essentially ‘sample’ across the sin curve.

The implementation requires subclassing the Hatcher class, by re-implementing the BaseHatcher.generateHatching and the BaseHatcher.hatch methods.

Unlike, the normal hatch vectors, the sinusoidal pattern has to be treated as a series of connected line segments, without any jumping. This requires using the ContourGeometry representation to efficiently store the discretised curve. As a result, the Hatcher.hatch method has to be re-implemented to take account of this.

The procedure builds upon previous methods to define customer behavior (see previous post). The first steps are to define a local coordinate system x' and y' for generating the individual sin curve. A sine curve y' = A \sin(k x') is generated to fill the region bounding box accordingly, given a frequency and amplitude parameter along x'.

The number of points used to discretise the sine curve is determined by \delta x. This needs to be chosen to suit the parameters for the periodicity and amplitude of the sine curve. A reasonable compromise is require as this will severely impact both the performance of clipping these curves, but also the overall file size of the build file generated.

dx = self._discretisation # num points per mm
numPoints = 2*bboxRadius * dx

x = np.arange(-bboxRadius, bboxRadius, hatchSpacing, dtype=np.float32).reshape(-1, 1)
hatches = x.copy()

"""
Generate the sinusoidal curve along the local coordinate system x' and y'. These will be later tiled and then
transformed across the entire coordinate space.
"""
xDash = np.linspace(-bboxRadius, bboxRadius, int(numPoints))
yDash = self._amplitude * np.sin(2.0*np.pi * self._frequency * xDash)

"""
We replicate and transform the sine curve along adjacent paths and transform along the y-direction
"""
y = np.tile(yDash, [x.shape[0], 1])
y += x

x = np.tile(xDash, [x.shape[0],1]).flatten()
y = y.ravel()

After generating single sine curve, numpy.tile is used to efficiently replicate the curve to fill the entire bounding box region. Each curve is then translated by an increment defined by x, to represent the effective hatch spacing or hatch distance.

The next important step is to define the sort order for scanning these. This is slightly different, in that the sort order is done per line segment used to discretise the curve. This is subtle, but very important because this ensures that the curves when clipped by the slice boundary are scanned in the same prescribed sequential order.

The increment of 1\times10^5 is used in order to potentially differentiate each curve later, if required.

# Seperate the z-order index per group
inc = np.arange(0, 10000*(xDash.shape[0]), 10000).astype(np.int64).reshape(-1,1)
zInc = np.tile(inc, [1,hatches.shape[0]]).flatten()
z += zInc

coords = np.hstack([x.reshape(-1, 1),
                    y.reshape(-1, 1),
                    z.reshape(-1, 1)])

Following the generation of these sinusoidal curves, a transformation matrix is applied accordingly, before these are clipped in the Hatcher.hatch method.

The next crucial difference, that has been implemented from PySLM version 0.3, is a new clipping method, BaseHatcher.clipContourLines. The following method is different from BaseHatcher.clipLines, in that clips ContourGeometry separately. This is important for keeping the scan vectors separate and in the correct order, which would be otherwise difficult to achieve. The clipped results are implicitly separated into contour geometry groups.

hatches = self.generateHatching(paths, self._hatchDistance, layerHatchAngle)

clippedPaths = self.clipContourLines(paths, hatches)

# Merge the lines together
if len(clippedPaths) > 0:
    for path in clippedPaths:
        clippedLines = np.vstack(path) 
        
        clippedLines = clippedLines[:,:2]
        contourGeom = ContourGeometry()

        contourGeom.coords = clippedLines.reshape(-1, 2)

        layer.geometry.append(contourGeom)

The next step is to sort the clipped paths into the right order. This is done by using the 1st value of 3^rd index column accordingly sorting using sorted with a lambda function.

"""
Sort the sinusoidal vectors based on the 1st coordinate's sort id (column 3). This only sorts individual paths
rather than the contours internally.            
"""
clippedPaths = sorted(clippedPaths, key=lambda x: x[0][2])

Now, the result of the sinusoidal scan strategy can be visualised below.

This approach currently is very intensive to generate during the clipping operation, due to the number of edges along each clipping operation. Using the previous techniques with the island scan strategy in a previous post, could be use to amorotise a lot of the cost of clipping.

Example Script

The script is available on github at examples/example_custom_sinusoidal_scanning.py