Real Time Simulation and Visualization of NC Milling Processes for

Inhomogeneous Materials on Low-End Graphics Hardware

Andreas Holger König and Eduard Gröller

f koenig j groellerg @cg.tuwien.ac.at

Institute of Computer Graphics, Vienna University of Technology

Karlsplatz 13/186/2, A-1040 Wien, Austria,
http://www.cg.tuwien.ac.at/home/

Abstract 1. Introduction

NC milling simulation using computer graphics tech-

Simulation and visualization of NC milling processes has
niques was proposed some years ago to verify NC programs
become an important step in computer aided manufactur-
rapidly and precisely.
ing. The usage of stock materials with specific locally vary-
Simulation systems exist for the two common types of
ing properties (like density, accuracy, color,. . . ) becomes
NC milling: 3-axis milling and 5-axis milling. Almost all
more and more important with new technologies emerging
industrial parts can be milled with three degrees of freedom
in the material industry. Our new approach, using volumet-
ric representation, has been adapted to this needs and copes
x; y; z . The values of x and y describe the movement of
the cutter in the plane defined by the top of the workbench,
with inhomogeneous material properties. Taking color as
whereas z is the height of the cutter above the workbench.
one possible material property, our approach enables the
The axis of the cutter is fixed to be vertical. Some appli-
visualization of milled wood or compound materials.
cations require the cutter axis to be tilted (5-axis milling).

Furthermore, our approach has been developed with the

This introduces two additional degrees of freedom:  and 
being the rotational tilt of the cutter axis. The generation of
usage of low-end graphics hardware in mind. The algo-
a tool path for 5-axis milling is quite sophisticated. There-
rithms have been optimized to ensure interactive update
fore, 5-axis milling is not as common as 3-axis milling.
rates even on standard personal computers without hard-
ware graphics acceleration. During the simulation of NC milling usually graphical
representations of the workpiece and the milling tool (the

Keywords: NC milling simulation, dexel approach, in- cutter) are displayed. The cutter can be observed moving

homogeneous material properties along the toolpath, which is defined by the NC program.
As the cutter is removing material, intersection calculations

1
are used to determine the removed parts of the workpiece. 1.1. Accurate approaches to NC milling
The workpiece is updated for every frame in the image se-
Kawashima [13] used a special geometric modelling
quence. The geometric representations of workpiece and
method called Graftree to speed up his solid modeling ap-
cutter are displayed on the screen using more or less so-
proach. Using CSG modelling, their approach allowed ac-
phisticated computer graphics techniques.
curate and precise representation of the workpiece and the
tool, while an octtree helped to decrease the number of ray-
The most important aspect in NC milling simulation is
intersection calculations in rendering. NC milling was sim-
error assessment. The workpiece is compared to the desired
ulated by using a Boolean difference operator. Another ap-
design part. The local deviation can be used to map the
proach to find the exact representation of the envelope of a
error into color hue, which can easily be interpreted by the
swept volume was chosen by Sourin [1]. He described the
NC programmer. Special constraints to error simulation like
cutting tool analytically with the use of procedurally im-
cutter wear-out are usually not taken into account. An accu-
plemented time-dependent defining functions. Boolean set
rate simulation of this phenomenon would require detailed
operations are defined by Rvachevs R-functions [5]. The
knowledge about material properties of the workpiece like
milling process is visualized by ray tracing the analytically
density and heat coefficents.
defined scene, which is a very time-consuming process.

Two major approaches to the simulation of NC milling

1.2. Approximate approaches to the simu-
processes have been developed: the exact, analytical ap-
lation of NC milling
proach and the approximate approach. The main problem
Van Hook [10] developed a real time shaded display of
with the accurate approach (mostly done in CSG) is its
a solid model being milled by a cutting tool which follows
computational expense. The cost of simulation is reported
an NC path. This approach utilized a dexel (depth element)
to be O(N 4 ) [9], where N is the number of tool move-
representation of the workpiece and cutter geometries. The
ments. A complex NC program might consist of ten thou-
data structure is a run length encoded version of a volu-
sand movements, making the computation intractable. In
metric data representation. An update rate of ten cutting
order to increase efficiency, a number of approximate simu-
operations per second was attained by using Boolean set
lation methods have been devised. The computational cost
operations on the one-dimensional dexels. The view point
of these methods simplifies to O(N ).
dependency of this approach was overcome by an exten-
sion of the method by Huang [12]. Huang also introduced
In the next section a detailed discussion of the previous the possibility of error assessment to Hooks method. Taka-
approaches is given. Then the dexel approach, being the fumi also used an extension of the z-buffer method (called
basis of our work is described. Our extensions and acceler- G-buffer) to simulate NC milling [8]. A completly differ-
ations to this approach are discussed. Special emphasis will ent approach was chosen by Jerard [2]. The design sur-
be given to inhomogeneous material properties, which are face is approximated by a polygonal mesh, where the sur-
accounted for in our approach. Finally some results will be face normals or other arbitrarily chosen vectors at the mesh
shown. points are intersected with a polygonal approximation of the
tool during simulation (therefore this approach is known as fore speeding up the method. A visualization technique
the lawn mowing approach). Yang [7] developed a method had to be developed for this low resolution, which produces
suitable for wire-EDM (wire cut electric discharge machin- polygonial output. Rendering of these polygon meshes can
ing), where cutting is only done at the sides of the work- utilize one of the standard 3D APIs like OpenGL.
piece. Glaeser [4] used differential geometric techniques to Second, the dexel structure itself has been simplified.
efficiently generate the swept volume of a moving cutter. Our investigations in the NC milling industry have shown,
Intersection calculations are done in data structure that is an that in 95 percent of all cases just simple forms of 3-axis
extension to the z-buffer data structure, called ?-buffer. milling are used. Most workpieces are flat plates, which are
For an overview of the techniques described see table 1. only machined from above. Our work is intended to give an
inexpensive solution for these most common machine ar-
2. The extended dexel approach to NC milling rangements. No internal holes or slots at the side faces of
the workpiece are produced by these simple kinds of NC
The dexel approach described in [10] and [12] is a spe- machining. Therefore, no multiple layers of dexels have to
cialized version of run length encoding the stock material. be stored. As no linked list data structures have to be used
A dexel (depth element) is a rectangular solid extending like in [10] and [12], our dexel structure is simply a two-
along one specified direction of the object to be represented. dimensional array. Detailed information on the dexel data
In [10] this direction is aligned with the viewing vector. sructure will be given in the following section.
The reolution of the dexel structure is fixed according to
the screen reolution. Therefore color values stored with
2.1. The dexel set
the dexels can be directly copied to the framebuffer for the
Let a dexel data set D be defined as the quadruplet
visualization of the dexel structure. The approach needs
(R; X; L; C ), with
enormous amounts of storage and interactive updates of the
milling scene are only possible on special graphics hard-
 R = (Rx ; Ry ) 2 N2 describes the resolution of the
volume data set in x and y dimension.
ware.
An independent coordinate system for the objects is used  X is a set of Rx  Ry dexels. The contents of a dexel
in [12]. Therefore the direction, in which the dexel extend, will be described later on.
can be arbitrary. As this direction is no longer aligned with
 L 2 R3 is the location of the volume data set in 3D
the viewing vector a special visualization technique has to
space. This point is usually the centre point of the grid
be employed. The creation of a shaded contour display, con-
plane containing the dexels.
sisting of lines, is discribed in this work. This visualization
technique is also too expensive for low end graphics hard-  C = (Cx ; Cy ; Cz ), with Cx ; Cy ; Cz 2 R3 , is the lo-
ware. cal coordinate system describing the orientation of the
dexel data set in 3D space.
For our purposes the dexel approach has been simplified
and adapted to the needs of 3-axis milling. First of all, we A dexel is a rectangular solid extending in the z direction of
are using a lower resolution than the original methods. This the volume data set (refer to figure 1). Let a dexel dex 2
reduces the amount of data which has to be processed, there- D:X be defined as the triplet (top; bottom;~n) where
Method workpiece cutter ea tool shape vp swept volume axes

Graftree[13] Octree CSG yes arbitrary no explicit as CSG 5

Rvachev functions analytical no arbitrary no analytically 5

method[1]

dexel approach[10] dexel no arbitrary yes calculated in image space 3/5

view point indepen- dexel yes arbitrary no calculated in image space 5

dent dexels[12]

G-buffer[8] ext. z-buffer yes arbitrary yes scanning of G-buffer 3

lawn mowing[2] mesh with normals polygons yes ball or cylinder no explicit as polygons 3/5

wire EDM[7] R-buffer (cylindrical z-buffer) yes wire no explicit as polygons 4

?-buffer[4] ?-buffer polygons yes arbitrary no scanning of ?-buffer 3/5

Table 1. Comparison of previous techniques. Abbrevations used for the table columns: ea. . . error
assessment, and vp. . . view point dependency.

y Cz bits, then 2m different height levels may be described

by top and bottom.

0 x
 ~n 2 R3 is the height gradient vector for the dexel. It
z is precalculated by the central difference method using
the top values of the dexel set.
L 0 1
Cx B dexi+1 ;j :top ? dex i?1;j :top C
B@ dexi;j+1 :top ? dexi;j?1 :top CCA
dexi;j :~n = B (1)
Cy R ?1
y
Rx with dexi;j 2 D:X . The use of ~n singnificantly speeds
up the simulation of milling free form surfaces, which
Figure 1. Dexel data set for a cube shaped ob- is described in section 2.4. As 3-axis machining of
ject free form surfaces is most common in industrial NC
milling, this work will focus on improvements in this
field.
 top is the top height-value of a dexel.
2.2. Workpiece and cutter representations
 bottom is the bottom height-value. The number of as dexel data sets
bits used to describe these values determines the res-
olution which may be achieved in z -direction. If Two different objects modeled with the dexel representa-
top; bottom 2 [0::2m ? 1] where m is the number of tion are needed for our approach of simulating NC milling:
a workpiece DW and a cutting tool DC . Cylindrical cutters (as well as the cylindrical shaft of ball-
end cutters) may be defined in a similar way.
2.2.1 The workpiece

In 95% of cases, NC milling is dealing with 3-axis milling

2.3. Cutting the workpiece
of flat plates. Therefore, we will describe the setup of the An instances of motion approach (described in [12])
workpiece for the case of a cube shaped solid. The follow- is used to move the cutting tool along the entire prede-
ing algorithm fills the workpiece dexel structure DW:X : fined tool path. At evenly spaced discrete locations, close
for all dexels dex[i,j] in DW.X enought that no dexels are skipped, the intersection opera-
if (i=0 or j=0 or i=DW.Rx or j=DW.Ry) tion is evaluated for cutter and workpiece.
dex[i,j].top=0
else
dex[i,j].top=height
dex[i,j].bottom=0

where height is the defined height value of the desired

block of material. Setting the height of border dexels to 0
assures, that the dexels on the edges of the dexel set produce
a surface during the visualization step. Figure 2. Possible intersections of cutter and
workpiece dexels
2.2.2 The cutter

Ball-end cutters are the most common tools in milling free The dexels of the cutting tool DC:X are trans-
form surfaces. Therefore the geometry of a ball-end cutter formed into the coordinate system of the workpiece
is described. The geometric definition of a sphere is used  i;j is within the
(DW:C; DW:L). If a transformed dexel dex
to describe the desired shape of the cutter DC . For every boundary of the workpiece dexel structure, then the nearest
location (i; j ) on the dexel grid of DC a ray is cast in direc- neighbour wx;y 2 DW:X is found. This discrete sampling
tion of DC:Cz . This ray is analytically intersected with the approach could lead to aliasing artefacts. A better solution
defining sphere. The intersection points define the geometry would be to take all dexels into account, which are covered
of the dexel dexi;j 2 DC:X . Three cases may occur: by the projected area of the cutter dexel. Cutting should

 Two intersection points ((i; j; zb ) and (i; j; zt )) are then be applied to all of these dexels, weighting the influ-

found. dexi;j :bottom is set to zb , the smaller one of ence of the cutting operation by the fraction of the dexel

the calculated z values. dexi;j :top is set to zt . area which is common to both the workpiece and cutter
dexel. This approach is computationally much more expen-
 One intersection point is found: (i; j; z ). Both
sive than chosing a single candidate dexel for cutting. As
dexi;j :bottom and dexi;j :top are set to z . the visual differences in results are neglegible (present, but
 No intersection point is found. Both dexi;j :bottom not visible), the inexpensive approach can be chosen.
and dexi;j :top are set to 0. The cutting operation is simplified to comparing the top
and bottom values of these two corresponding dexels. The Converting the 3-axis milled dexel structure into a
following cases (depicted in figure 2) may occur: surface representation is straightforward. The top loca-
tions of four neighbouring dexels dexi;j :top, dexi+1;j :top,
 wx;y :top < dex
 i;j :bottom: The cutter dexel is above
dexi+1;j+1 :top, dexi;j+1 :top determine the vertices of two
the workpiece dexel. No intersection is found. (figure
triangles. Evaluting this for all dexels gives a representation
2a)
of the top surface without holes (see figure 3).

 wx;y :bottom < dex

 i;j :bottom < wx;y :top < Non-trivial cases occur, when top and bottom locations of a

 i;j :top:
dex The cutter removes the top part of the
workpiece dexel. wx;y :top is set to  i;j :bottom.
dex Cy
surface
Cz
(figure 2b)
surface

 wx;y :bottom < dex

 i;j :top < wx;y :top and
 i;j :bottom < wx;y :bottom:
dex The cutter removes
the bottom part of the workpiece dexel. wx;y :bottom X (i,j)
Cx
X
(i,j) Cx

 i;j :top. (figure 2c)

is set to dex
Figure 3. Connecting top locations for sur-
  i;j :top < wx;y :top and wx;y :bottom <
dex face generation
 i;j :bottom: The cutter removes an inner part of
dex
the workpiece dexel (figure 2d). This would produce
two dexels for wx;y . We are assuming, that this case
dexel have the same value (see figure 4). It has to be defined,
can not be produced in our 3-axis milling simulation.
whether such dexels are still contributing to the border of
Changes to the workpiece dexel are ignored.
the hole (like in figure 4a the leftmost and rightmost dex-

 wx;y :top < dex

 i;j :top and  i;j :bottom <
dex els) or if they are already part of a hole (the dexel produces

wx;y :bottom: The cutter removes the workpiece dexel no surface, like in figure 4b the leftmost and rightmost dex-

completely (figure 2e). wx;y :top and wx;y :bottom are els). When only the top surface of the dexel structure has

both set to zero. to be created, both alternatives depicted in figure 4a and 4b

could be used. Dexels with both top and bottom value set
2.4. Visualization to zero could be treated as lying on the border of the hole
(still within material, figure 4a) or as lying in the hole (no
Several methods for the visualization of dexel data sets material, figure 4b).
do already exist. Van Hook [10] displays the data set by di- When the bottom surface of the dexels structure is also cre-
rectly copying it to the video frame buffer, whereas Huang ated, treating these dexels as border dexels (producing sur-
[12] uses a contour display method to render the dexel struc- face) leads to incorrect results. Refer to figure 4c and 4d
ture. As we are using a lower resolution than in the above for a comparison. The next section will focus on a special
approaches, we had to develop a new visualization tech- algorithm for the creation of visually correct surface repre-
nique which produces polgonial output. sentations for such non-trivial dexels.
 Cz Cz

surface surface
0 1 2 3

Xt = Xb = 0 vertical horizontal
Xt = Xb polygon polygon
Cx Cx
X (i,j) X (i,j)
a b
Cz Cz
Figure 5. Possible cases of border/hole dis-
surface surface
tribution among neighbouring dexels

are classified as ”material“ (figure 5(1)). A vertical

(i,j) Cx (i,j) Cx
X X polygon is created between them to close the side face
c d
of the dexel structure.

Figure 4. The non trivial case of connecting  Three dexels are classified as ”material“ (figure 5(2)).
top and bottom surfaces One horizontal polygon is created to close the top sur-
face, one vertical polygon along the diagonal is created
to close the side face.
2.4.1 Obtaining visually correct vertical edges

An adapted version of the well known marching cubes al-

 All four dexels are classified as ”material“ (figure
5(3)). Two horizontal polygons are created to close
gorithm [11] is employed to treat all possible cases in a vi-
the top surface.
sually correct manner. Four neighbouring dexels are used
to create a surface part. First they are classified, whether The use of the vertical polygons along the diagonal reduces
both top and bottom values are zero (”holes“) or not (”ma- the influence of object-space aliasing. Instead of having a
terial“). This knowledge is used to create an index into a set of vertical polygons where two adjacent faces are either
table of all possible cases. As with the original marching coplanar or meet perpendicularly, now also a angle of 45 de-
cubes algorithm, all possible cases can be derived from a grees is possible. Nevertheless, aliasing is still visible. The
basic set by rotation and inversion. See figure 5 for the setup next section deals with a method to overcome this problem.
of basic possibilities. Some cases (like two holes on a di-
agonal) can not be produced by rotation or inversion from 2.4.2 Antialiasing
the basic set. These cases dissolve automatically into basic
Aliasing occurs at the side faces of the dexel structure.
cases, when successive rows of dexels are processed. Four
The top surface polygons are shaded with the precalculated
different cases may occur (refer to figure 5):
height gradient normal vectors, therefore aliasing artifacts

 All four dexels are classified as ”hole“ (figure 5(0)).

are rare. As the side faces are always vertical, their gradient
vectors lose one degree of freedom and aliasing is intro-
No polygons are created at all.
duced. To overcome this problem, a standard antialiasing
 Two dexels sharing an edge in the square of interest technique, blurring, was adapted to be used for the normal
vectors. For 2D raster images blurring provides antialiasing 2.5. Error assessment
by averaging the weighted color values of adjacent pixels.
This convolution operation averages the color values and Error assessment can easily be achieved by comparing

high frequency changes are eliminated from the image. We the dexels of the workpiece DW:X with the dexels of the
are using averaging on 3D vectors: the normal vectors of design part DD:X . Local error assessment compares single

neighbouring dexels and the normal vector of the dexel are dexels whereas global assessment compares all the dexels.

weighted and summed up to get the blurred normal vector Overcutting and undercutting may occur (see figure 8).

for the dexel. See figure 6 for a graphical representation Overcutting occurs at regions, where already too much ma-

of the process. When these antialiased normal vectors are terial has been removed. Undercutting means, that there is
still material in excess at some location.

Figure 6. Antialiasing by blurring the normal

vectors

used for shading, the lighting calculation for the polygons

is modified. The change of color between adjacent poly-
Figure 8. Comparison of workpiece and de-
gons is less dramatic than when using the exact normals of
sign part dexels gives information on over-
the side face polygons. The geometric representation is of
and undercutting
course unaltered and still precise (with respect to the under-
lying resolution). See figure 7 for a comparison.

2.5.1 Local error assessment

Two strategies have been investigated to visualize the local

error.

 A rough overview about regions, where the tool path

Figure 7. shading without antialiasing (left) has to be reprogrammed, is gained, when design part
and with antialiasing (right) and workpiece are rendered simultaneously. If the
workpiece is drawn transparently, even more informa-
tion about the undercutting error is visible. For an ex-
 ample see figure 9. If overcutting has to be visualized, fore homogenity throughout the whole block of stock ma-
the design part has to be drawn transparently. terial can no longer be assumed. In our new approach to
NC milling simulation, this fact is taken into account. We
are investigating the possibilities when dealing with mate-
rial properties in the next sections. Assuming that material
properties may vary in different regions of the workpiece,
various fields of application are possible. Some examples
Figure 9. Error assessment by drawing the
will be given: If cutter speed limits can be defined for differ-
workpiece transparently over design part
ent regions of the workpiece, the automated tool path gen-
eration can be further optimized in means of time. Parts of
 A more sophisticated stategy is to encode the lo- the workpiece with higher density have to be milled slower
cal error in the surface representation of the work- in order to not wear out the cutter too early. Different ma-
piece by the use of color. The difference between terials could also be milled with different rotation speeds
workpiece dexel and design part dexel (dexi;j :top ? of the cutting tool. Some other machine parameters could
dpx;y :top; dexi;j 2 DW:X; dpx;y 2 DD:X ) is also be controlled with varying material properties. For in-
mapped via a transfer function into a color value. This stance during milling of a compound material consisting of
value is used for the vertices of the top surface part metal and wood the cooling liquid of the NC machine has
created by this dexel. One possible transfer function to be stopped during wood milling, in order to prevent the
could map overcutting to red, whereas undercutting is wood from soaking up the polluted liquid. Another appli-
mapped to green. If no error occurs, the vertices will cation could be the definition of varying error tolerances for
be assigned the color white. For an example of this different parts of the workpiece. Automated tool path gen-
visualization method see figure 10. eration could generate high accuracy movements for, e.g.,
bearings and surfaces of contact, whereas not so important
2.5.2 Global error assessment
parts could be treated with higher error tolerance limits.
Global error assessment is done by summing up the local
errors.
X
E= (dexi;j :top ? wi;j :top)2 (2)
i;j
where dexi;j 2 DW:X; wi;j 2 DD:X and E is the global
Color as a material property must be taken into account
error. Information about the global error can be given as an
absolute value (E ) or as a relative value (Erel ):
during the simulation of NC milling. Milling a piece of

Erel = P wE :top
wood or a block of compound material will not give the
(3)
i;j i;j same visual impression as milling a block of grey alu-

2.6. Material properties minium. Furthermore, color is a simple and intuitive exam-
ple of a varying property, which may change with high fre-
Nowadays often special metal alloyes, compound and quency. The next sections will focus on the usage of color
other syntethic materials are used for NC milling. There- for the simulation of NC milling processes.
 Figure 10. Error assessment by color coding

2.7. Color well behaved polygon

subdivided
Different methods are possible to simulate the use of bad polygons
behaved
polygon
color. Color can be predefined by the user or it can be as-
sociated with special local properties and created on the fly
during isosurfacing. Predefining colors can also be done by
Figure 11. Bad behaved polygons have to be
simply painting them as 2D images or specifying them with
converted into well behave ones
the use of analytical functions.

be the texture function for dexel set D with

8
2.7.1 Procedural defined 3D textures < black z > half the block height
T (x; y; z ) = : (4)
Procedural defined textures can be evaluated for points in
yellow otherwise
3D space. The procedural texture function is evaluated for The color value returned for point (x; y; z ) (in dexel coor-
every point used to build the surface representation. We are dinates) only depends on the z value. If it is within the top

not doing real texture mapping for all points in the interior part of the workpiece, the texture function will evaluate to

of a triangle. By only assigning color values to the vertices black, whereas it will be yellow if the desired point is be-
and using smooth shading, large polygons become clearly low half the height of the workpiece dexel set. An image

visible for the lack of color change in their interior. There- of a workpiece milled with this function is shown in figure

fore, large polygons (and especially the extremely thin and 12. More sophisticated procedural texturing functions in-

long (bad behaved) polygons of the side surfaces) have to clude the stochastic textures described in [3]. For images

be split into a set of smaller (well behaved) polygons. Refer rendered with these textures see figure 14 to 16.

to figure 11 for a graphical representation of bad and well

2.7.2 Physically based Textures
behaved polygons. One simple example for a procedural
texture could describe the block of stock material as con- Another possibility to define a texture for a dexel set is
sisting of two layers of differently colored plastic: Let DT to interpret local geometric properties of the dexels. For
 Figure 12. Design part (left) milled from two layered compound material (right)

instance, the local derivative of the dexel height could be ated with this prototype implementation.
mapped into color values. High frequency height changes
For high quality color versions of the images presented
could be emphasized, therefore accentuating edges. Other
here, please see our web page at
possibilities include the highlighting of regions where the
dexel set is very thin, e.g., the difference between dexel
height and bottom values is small. This knowledge is very http://www.cg.tuwien.ac.at
important for the NC programmer in order to treat fade-out /research/vis/VolVis/NC/nc.html
regions carefully, because they tend to produce cracks.

The workpiece shown in figure 13 has been modeled

3. Results
with the aid of a scanned photograph of the instrument panel
In order to prove the usability of the concepts described of a Honda Fireblade motorbike.
in the previous sections, a software prototype was imple-
An analytically defined workpiece (with milling still in
mented. As the method should be suitable for low-end
progress) is shown in figure 14. It has been milled from a
graphics hardware, an IBM compatible 200 MHz Pentium
block of stock material consisting of white and blue marble.
Pro computer was chosen as the target platform. No 3D
Figure 15 shows another geometry milled from two lay-
graphics accelerator boards were used. The prototype was
ers of marble.
implemented using Microsoft Visual C++ (ver. 4.2) under
Microsoft Windows NT 4.0. The rendering of the polygo- Figures 16 shows a replication of a tsuba, which is a
nial representations of the dexel structures was done with part of an ancient japanese sword. The workpiece has been
OpenGL. All images presented in the following were cre- milled from bronze.
 instrument panel (figure 13) smiley (figure 12) free form surface (figure 14)

resolution 80  40 128  128 256  256

cutter size 99 12  12 66
cutting operations per second 30 28 32
complete surface extraction 20 sec 1 min 20 sec 2 min 30 sec
local update of surface 100 msec 120 msec 80 msec
instances of motion 6400 16512 65792
simulation time 10 min 40 sec 33 min 4 sec 1 h 27 min 7 sec
cutting only 3 min 30 sec 9 min 42 sec 34 min 16 sec

Table 2. Comparision of processing times for three different workpieces

3.1. Statistics for the different workpieces. The tool path has been gen-
erated automatically. The geometry of the design part was
Table 2 gives a comparison of the processing times for
scanned for height in x-axis parallel stripes. When the scan-
three different workpieces. This test runs were investigated
ning process reached the edge of the design part, the in-
with our prototype implementation on a 200MHz Pentium
struction to lift the cutter was added to the NC programm.
Pro computer without any hardware graphics acceleration.
The cutter was instructed to return to the other edge without
The first row of table 2 gives the resolution of the dexel
milling, where scanning of the design part was continued.
data sets for the workpiece in x and y direction.
This kind of machining is called zig-milling [6]. As in our
The second row describes the resolution of the cutting
implementation the instances of motion calculation could
tool. All workpieces presented here were created by ball
not be turned off for the moving cutter, intersection calcu-
shaped cutters.
lations were also made on the parts of the tool path where
The third row gives the number of cutting operations,
the cutter just returned to the other edge of the design part
which are possible per second.
without milling. Therefore the number of instances of mo-
The fourth row of table 2 gives the time neccessary for
tion (given in the sixth row of table 2) is in general two times
the visualization of the complete dexel data structure of the
the minimal number of instances neccessary to produce the
workpiece. This has usually only to be done once at the
workpiece.
beginning of the simulation process. During the simulation,
when the cutter moves over the workpiece, only the local The last two rows of table 2 show the simulation time
part of the scene has to be recreated which is affected by for milling the three different workpieces. The row before
the cutter motion. The time for this local update operation is the last gives the overall simulation time for the different
given in the fifth row. As it can be seen by the small amounts workpieces. For the measurements given in the last row
of update time, a continuous real time display of about 10 (”cutting only“), the animated visualization of the milling
frames per second is possible. A smooth animation of the process has been turned off. In this case just the cutting
moving cutter can be observed. calculations were made without the continuous update of
The sixth row gives the number of instances of motion the graphical representation of the milling scenario.
4. Conclusions tute of Computer Graphics, University of Technology,
Vienna., 1997.
A new approach to the simulation of NC milling pro-
[5] Rvachev V. L. Methods of Logic Algebra in Mathe-
cesses was created. The well known dexel approach uses
matical Physics. Naukova Dumka Publishers, Kiev,
a high resolution data structure, which requires enormous
1974.
amounts of processing power. The dexel approach has been
adapted to the limited performance capabilities of low-end [6] Held M. A geometry-based investigation of the tool
graphics hardware by using a lower resolution of the data path generation for zigzag pocket machining. The Vi-
structure. Therefore a special visualization technique had sual Computer, 7:296–308, 1991.
to be developed, which produces renderings of low reso-
[7] Yang M. and Lee E. NC verfication for wire-EDM
lution dexel structures. Even on low end IBM compatible
using an r-map. Computer Aided Design, 28(9):733–
platforms, update rates are interactive.
740, 1996.
Another aim of this work was the integration of inho-
mogeneous material properties. As the dexel approach is [8] Takafumi S. and Takahashi T. NC machining with
a volumetric representation of data, various different data the G-buffer method. Computer Graphics, 25(4):207–
values can be stored and processed along with the geomet- 216, 1991.
ric information of the workpiece. Color was chosen as an
[9] Chappel I. T. The use of vectors to simulate material
intuitive example of a locally varying property. Using pro-
removed by numerically controlled milling. Computer
cedurally defined 3D textures, visually realistic results were
Aided Design, 15(3):156–158, 1983.
obtained.
[10] Hook T. van. Real time shaded NC milling display.

References Computer Graphics, 20(4):15–20, 1986.

[11] Lorensen WE. and Cline HE. Marching cubes: a high

[1] Sourin I. A. and Pasko A. A. Function representation
resolution 3D surface reconstruction algorithm. Com-
for sweeping by a moving solid. IEEE Transactions
puter Graphics, 21:163–169, 1987.
on Visualization and Computer Graphics, 2(2):11–18,
1996. [12] Huang Y. and Oliver J. H. NC milling error assessment
and tool path correction. Computer Graphics Pro-
[2] Jerard R. B., Hussaini S. Z., Drysdale R. L., and
ceedings, Conference Proceedings July 24-19, 1994,
Schaudt B. Approximate methods for simulation and
(Proc. SIGGRAPH ’94), pages 287–294, 1994.
verification of numerically controlled machining pro-
grams. The Visual Computer, 5:329–348, 1989. [13] Kawashima Y., Itoh K., Ishida T., Nonaka S., and Ejiri
K. A flexible quantitative method for NC machining
[3] Ebert D. Texturing and Modeling: A Procedural Ap-
verification using a space-division based solid model.
proach. AP professional, Cambridge, 1996.
Visual Computer, 7:149–157, 1991.
[4] Glaeser G. and Gröller E. Efficient volume generation
during the simulation of NC-milling. TR at the Insti-
 Figure 13. Instrument panel, part of a motorbike

Figure 14. Free form surface milled from marble, cutter in red
Figure 15. Another example for a workpiece milled from two layers of marble

Figure 16. Tsuba milled from bronze