After accuracy and placement quality (defects per million), a pick-and-place machine’s output is perhaps its most important parameter. It determines the number of boards that can come from a line, and thus the line’s return on investment.
The output of a pick-and-place machine depends strongly on the application, though. The number of components actually placed per hour will normally depend on the board size, the type of components placed, and even the accuracy needed to place them. Different manufacturers used to specify the outputs under their own conditions, and the temptation was, of course, to show the manufacturer’s own machine in the best light. That was misleading for purchasers, since the figure was not necessarily related to the number of components their lines would actually place.
To clarify matters, an IPC SMEMA committee in 2002 wrote IPC-9850, Surface Mount Placement Equipment Characterization. This standard establishes measurement procedures for specifying, evaluating and verifying surface mount placement equipment. The standard details how measurements must be made, and emphasizes consistency as the best way to compare them.
IPC-9850 has become a basic industry reference, but doesn’t tell the whole story. Several pick-and-place machine manufacturers, for example, currently claim the industry’s fastest placement speeds on the basis of the IPC reference speed. This usually comes, however, from selecting the most favorable IPC-9850 conditions. Speeds can be measured for a simple matrix of one type of component, for example, with “gang pick” operations, giving artificially high figures. These theoretical maximum speeds do not necessarily relate to the actual output in a particular application.
Placement must be accurate and repeatable at full output – otherwise a line becomes a fast way of producing scrap. Lines must have the lowest possible DPM figures to reduce or eliminate the need for rework (notwithstanding rework becomes more difficult as component sizes shrink).
How speed in practice compares to the IPC-9850 speed will depend on placement machine design and the application it is used in. If the pick-and-place machine design is good enough, machines can actually run above IPC-9850 speeds in an application with no loss in first-pass yield. As such, to avoid misunderstandings, purchasers should look further than IPC-9850. They need to compare accuracy and DPM specifications at the IPC reference speed, and should also ask for reference speeds for their specific application, rather than just relying on raw IPC figures.
The first basic figure of merit for a pick-and-place machine is the precision or accuracy with which it places components. Most commonly, this is defined by x, y and phi. These are the two-dimensional variables of placement deviation between component and footprint bodies, related to the substrate XY Cartesian frame (Figure 1).
Pick-and-place machine users are interested in the overlap between component terminations (lead, end-cap, ball or column) and footprint lands (pads) on the substrate. A good overlap means reliable connections. The required coverage depends on the selected connecting process (solder paste, conductive glues and so on) and footprint layouts. Selecting lands instead of solder paste isolates SMT placement machine performance from that of other flowline machines such as screen printers.
IPC-9850, therefore, introduced the required MLTE (Maximum Lead Tip Error) for good termination-to-land coverage. MLTE values give the maximum placement deviation acceptable between component termination and footprint land for all component terminations. Values depend on components, board layouts, footprint and other dimensions, and on manufacturing tolerances.
In practice, there will be a correlated MLTE placement accuracy requirement between the x and y directions for most of the interconnection processes. The combination of the x and y displacement must guarantee a specific overlap area (Figures 1 and 2). This placement “process window for joining and connecting” in the MLTE x/y area is expected to be circular or elliptical.
In IPC-9850, the coverage for leaded and end-cap components is provided as the percentage of the termination width placed on the land. That is determined by the MLTE in the direction perpendicular to the termination length (one direction only: x or y) due to the joint x (or y) and phi errors.
For ball or column arrays, though, the coverage defined by IPC-9850 is the percentage of the round termination area in contact with the land. Instead of the MLTE, IPC-9850 here specifies the maximum ball-to-land error (MBTL) due to joint x, y and phi errors.
IPC-9850 explicitly specifies the required MLTE/MBTL values for SOIC16, QFP100, QFP208 and BGA228 to guarantee a minimum of 50% (accuracy class 1, 2), and 75% (the more accurate class 3) termination-to-land coverage.
Another placement process window can account for the space allowed between footprints on the boards. The “spacing process window” is typically square. Placing within a window protects from collision with neighboring lands and placed components.
Most PCB layouts have a smaller placement process window for connecting than for spacing. However, typical high-miniaturization applications use small, closely spaced components. Assembling 0201 SMDs for minimum spacing can therefore have a more significant placement process window than that for connecting. It is always, however, the intersection of connecting and spacing placement process windows that will define the placement window required from the SMT process.
R0201 components with 100 µm spacing (400 µm pitch and 300 µm component width), for example, have lower MLTE limits for spacing than for connecting at 50% coverage. Figure 3 shows the intersection of the spacing and connecting process windows. Because spacing here is more critical than connecting, the placement machine must be within the squared MLTE process window.
Technical specification limits. Placement machine manufacturers commit to place SMDs within some tolerance limits according to the technical capability reported in their specification sheets. IPC-9850 calls these limits the technical Specification Limits (SL).
Noncorrelated x and y SLs are typical, to check x and y capability independently. SL_x and SL_y appear as four lines, two in the x direction and two in the y direction, in placement deviation plots (Figure 4), creating a square “technical process window.”
A circular SL shape would actually have been more appropriate, fitting into both measured Gaussian PDF (Probability Distribution Function) shapes and the required MLTE process window from the connecting process itself. Whichever shape is chosen, though, this area must always fit in the process windows both for connecting and spacing.
Quantifying accuracy. Placement deviations can be classified in two groups: deterministic and stochastic errors.
Deterministic errors have the same values irrespective of component type, footprint position and substrate. Most deterministic errors result from calibration errors. They often are called module errors because most are module-specific. A well-known cause is the mechanical interfacing of modules in the SMT machine after offline calibration. Module wear and thermal expansion of different parts of the SMT placer can introduce a deterministic error in a test batch, which may vary smoothly with time when testing is repeated.
Stochastic errors have different values for each component placement, and are often called process errors because most of them are specific to the physics of the placement process. They result from random variations in the placement process due to a number of factors. Those include mechatronic errors (robot friction called “reproducibility errors,” and vibrations called “dynamical errors,” machine wear (dust, vibration loosening of some machine parts), measurement errors inherent to sensor physics (lighting, digitizing, and so on), and materials manufacturing tolerances (substrate dimensions, warp and artwork deposits, and components).
For statistical purposes, placement deviation values are often grouped for different components placed. The deterministic errors cause the offset (average or mean measured placement deviations, µ). The stochastic errors generate the standard deviation Σ (sigma or spread) of measured data.
The statistical population of placement deviations is typically a Gaussian distribution. Most values are found around the average, while the probability of measuring a certain deviation decreases with the distance from the average (Figure 4). When x and y are independent variables, the probability of measuring both a certain value x and a certain value y is the product of the respective probabilities:
P(x y) = P(x).P(y)
The 2-D plot of the placement error PDF, characterized by different spreads in x and y, typically consists of concentric ellipses with axes parallel to the X and Y axes. The geometry becomes circular for equal spreads in x and y.
Sometimes x and y placement deviations have correlated values. One typical cause is angular uncertainty – transforming coordinates between two machine XY Cartesian frames by a Θz (phi) rotation. Another is scaling uncertainty – from varying optical amplification in a vision system due to spread in component thicknesses and z-control repeatability.
The correlation value is highly dependent on the positions of the objects within the frames of interest. Correlation may totally disappear when these positions are randomly spread in that frame. Where <x,y> correlation exists, the PDF assumes a 2-D elliptical geometry with axes rotated with respect to the X and Y axes (Figure 5).
To simplify further discussion, measured PDFs will be limited to 2-D circular shapes.
MLTE quantification. IPC-9850 indirectly quantifies the pick-and-place overlap between component terminations and substrate lands using the measured MLTE machine performance. The measured MLTE is the maximum deviation measured for that termination at maximum distance from the component body center, as result of combined x, y and angular (phi) inaccuracy.
Statistics in x, y and phi are used to calculate a MLTE offset and standard deviation.
Machine capability (Cpk). Machine capability indicates how well and controlled the measured pick-and-place process meets the technical, connection or spacing process window Specification Limits. It uses process capability indices Cp and Cpk. Values of these capability indices are calculated from the placement deviation measured average and standard deviation (Σ), shown in Figure 6.
It is the placement accuracy at the machine’s full output that is important to fix the capability with respect to the selected SLs.
Processes that are not capable are often unstable, with standard deviations too large (Cp<1.33) and/or averages that are too large (Cpk<1.33). Deterministic errors can typically be improved by better calibration methods and/or closed control loops to center the process capability (average = 0; Cpk>1.33). The stochastic errors of a process are specific to the chosen machine concept and design. Processes with a large spread (Cp<1.33) are difficult to improve inline. The best machine designs are those with a small stochastic/deterministic error ratio, placing the component comfortably in the center of the technical specification area.
Capability index vs. reject levels. Each capability index value has an associated reject level: the relative number of placement actions falling outside the SLs. Each placement deviation in the (assumed) Gaussian distribution tail exceeding SLs gives a direct penalty.
When x and y requirements are considered not to be correlated (rectangular process window), any x value exceeding the SL limits must be rejected, regardless of the related y value measured. Placement processes with a relatively large average placement deviation in x (or y) have PDFs that are not centered about zero. In this case, all “inaccurate” placements leading to rejects lie at one tail of the PDF (unilateral failure rate). Centered PDFs (Figure 7) have rejects at the two tails of the PDFs (bilateral failure rate).
For users, it is the first-pass yield that is most important. This depends on how many rejects will come from the line. Reject volumes are typically quantified in parts per million (ppm). These ppm values are estimated from the integration of the PDFs tail parts outside the process SLs, shaded red in Figure 7.
Acceptance tests have shown placement processes from different machines to perform very differently with the same technical specification limits. Reject levels vary between 100 and thousands of ppm, with the largest ppm values indicating worst machine reliability.
For most component types, the connection and spacing MLTE SLs have larger value than the technical SLs. Production, therefore, has higher Cpk values and lower reject levels than acceptance tests. Below 10 ppm rejects are really needed for high first-pass yield.
A 100 µm, 4-sigma machine specification tells the user the SMT placer can perform within the technical SL of 100 µm with a capability index Cpk of 1.33, with a maximum 63 ppm reject level. Increasing the Cpk values implies decreasing the reject levels. A Cpk of 2.0 translates into a 0.002 ppm reject level. First-pass yield values are given in Figure 8 for different Cpk values.
For a circular specification area, a good connecting process only can be guaranteed when each combination (x,y) exceeding the circular specification area limits is rejected. This approach would result in ppm values differing from those in Table 1, but more realistic.
Equipment manufacturers dream of step reductions in technical SLs (stricter mounting accuracy) together with step increases in related reliability level (lower ppm level, better fit of data in specification area, higher Cpk value). This is how to reduce assembly costs and take advantage of new connection technologies and even new market sectors.
However, even comparing placement machines is not straightforward because, if nothing else, the values for placement errors chiefly depend on the statistical experimental design chosen (Figure 9). The accuracy that is measured will be influenced by several factors. They include the:
IPC-9850 solves the problem of comparing machines by using standard materials and a standard statistical experimental setup. IPC-9850 specifies glass boards (placement verification panel, PVP) and glass slugs or dummy components. That gives the best possible base for verifying machine accuracy.
Using glass materials permits an optical coordinate measurement machine (CMM) with backlight to be used to measure placement. They also help make placement machine specifications independent of material manufacturing tolerances. Importantly, the responsibility for the quality of the materials to assemble lies with the user, not with the machine OEM.
IPC-9850 also specifies the dimensions and fiducial marks of the PVP. Four fiducials have specified locations with spacing accuracy traceable to a NIST certification standard. Extra fiducials may be required to characterize assembly concepts using artwork recognition. Here, artwork features can be selected anywhere on the substrate to align the board in the machine. The standard also defines the number of parts to mount for each panel and the number of panels to be built, together with the geometrical and statistical processing formulae to use.
All this also helps to ensure manufacturers specify the output of their machines consistently. There will, however, be differences between IPC-9850 speeds and the actual speed obtained by a line for a specific application.
The first major factor affecting IPC-9850 speeds is the board size. For a mobile phone board and modules, typical board dimensions are smaller than the 200 x 200 mm reference. For a mobile phone, for example, they are around 220 x 100 mm (L x W) with about 1200 components for a four-up panel. This will affect the travel distance, board transport and board alignment times.
So, on an actual mobile phone manufacturing line, the placement heads will not travel far. With 1200 components per board instead of IPC-9850’s 400 components, the influence of board transport and alignment will be less. Both these factors permit higher outputs in an application than the IPC-9850 reference would suggest. These can be small differences, but they can be around 20% of the overall pick-and-place cycle, and the difference between several thousand components per hour. The IPC-9850 output of one studied machine with 20 placement robots, for instance, is around 121,000 components/hr., while the actual output in a mobile phone line has been calculated at 148,000.
The difference in board sizes affects all placement machines equally. However, a major loophole in IPC-9850 could be used to boost the IPC-9850 output to a figure unreachable in an actual application. IPC-9850 speeds are measured by placing a simple matrix of components (for example, 80 SOIC-16s or 400 0603 capacitors) on a 200 x 200 mm substrate. However, the standard says nothing about the electrical value of these 400 capacitors (for example). That permits sequential placement machines to “gang pick” (simultaneous pick by multiple placement heads). Gang picking artificially inflates the performance figure, as it can virtually never be used in a real-world application.
The flaws of IPC-9850 should be considered before making a placement machine determination. The secret of reaching high first-pass yields at full output is to make assembly processes stable (in control), and then continuously work on them to reduce variation and so improve accuracy and repeatability. If a machine can place components well within customer specification limits, this will equate to consistently reliable placements.
Sjef van Gastel is manager advanced development at AssemblÃ©on Netherlands B.V. (assembleon.com); firstname.lastname@example.org.