Practical examples of parallel photonics alignment automation

Scott Jordan, senior director, NanoAutomation technologies, Physik Instrumente (PI)

Fast multi-channel photonics alignment (FMPA) technology, developed by Physik Instrumente (PI), is a set of firmware-level commands built into extremely high-performance digital nanopositioning and hexapod controllers. These commands allow fast coupling optimisation between photonic and other optical devices and assemblies, including optimisation across multiple degrees-of-freedom (DoF), inputs and outputs, elements and channels. Importantly, these optimisations can often be performed in parallel, even if the individual optimisations interact. Examples where significant process savings can be achieved span the spectrum from multichannel silicon photonics (SiPh) devices to light detection and ranging (LiDAR) sensors to smartphone camera assemblies.

Serial versus parallel alignments

For example, in the short waveguides increasingly utilised in SiPh devices, the input and output couplings can steer each other. As one side is optimised, the other shifts slightly and needs re-optimisation. Formerly, this necessitated a time-consuming, serial sequence of back-and-forth adjustments of the input, then the output, repeating until a global consensus alignment was eventually achieved. Similarly, when optimising an angle, the transverse alignment would be impacted and would conventionally need to be re-optimised, again in a time-consuming serial loop.

But with FMPA, these interacting alignments can often be optimised simultaneously, in parallel. This allows a global consensus alignment to be achieved in one go. Tracking and continuous optimisation of all the alignments is also possible in many circumstances, allowing compensation of drift, curing stresses and so on.

The results are much higher production throughput and often dramatically lower costs. As devices become more complex and precise, and as their production and test requirements grow more demanding, this parallelism is increasingly critical to process economics.

Look for the loops

Fully exploiting this capability for maximum overall cost savings can require some different thinking than what one might be used to with classical alignment hardware. In general, one looks for loops of sequential alignments, for which simultaneous optimisations can usually be substituted. This article reviews a few sample applications and discusses implementation issues to illustrate how this remarkable capability can be utilised to maximise productivity in test and packaging.

Background of FMPA operation

The device alignment should be broken down into discrete alignment processes. For example, probing a waveguide with one input and one output using lensed fibres typically involves the following four alignment processes:

1. Transverse optimisation routine, input.
2. Transverse optimisation routine, output.
3. Z optimisation routine, input (beam waist seek).
4. Z optimisation routine, output (beam waist seek).

If the device has one or more additional inputs or outputs, the following two alignment processes are added as necessary:

1. Theta-Z optimisation routine, input.
2. Theta-Z optimisation routine, output.

If the device requires optimisation in theta-X and theta-Y, the following two alignment processes are added:

1. Gimbaling optimisation routine, input.
2. Gimbaling optimisation routine, output.

And so on. Dividing the overall alignment task into these sub-processes is key to identifying which processes can be performed simultaneously.

With FMPA, one takes the list of alignment routines and defines these directly into the controller. This only needs to be done once (and can be changed or updated at any time). Once a routine is defined, it can be executed repeatedly. More than one routine can be executed at a time, this is the parallelism.

Defining a process means instructing the controller which axes are involved in the process, which analogue input presents the quantity to be optimised (optical power, modulation transfer function (MTF)...), and various process options. Each process is given a name; the numbers in the list just made are perfect for that.

Routines are executed by issuing the fast routine start (FRS) command. Referring to the list just constructed, FRS 1 would commence the transverse optimisation on the input, FRS 2 would commence the transverse optimisation on the output and FRS 1 2 would do both at once.     

Types of alignment routines

For each side of the device, independent alignment engine hardware is, of course, necessary. Any number of alignment engines can be used; most common configurations utilise one or two, but three or more will be increasingly common as SiPh technology matures.

Most often, each alignment engine is constructed of a multi-axis, long-travel assembly and a shorter-travel, high-speed, high-resolution piezoelectric multi-axis nanopositioning stage. The modularity of the approach is a key benefit. Some applications do not require the long-travel mechanism; some applications do not require the speed, resolution or continuous tracking capability of the nanopositioning stage. In any case, all FMPA algorithms and processes are virtually identical regardless of the type of motion system involved; only the capabilities will differ.

Long-travel options

For situations requiring no angular optimisations or array alignments, a stack of linear stages is sufficient. Otherwise, a hexapod is requiredoptimal, not only in situations where full six DoF positioning and optimisation are necessary, but also in simpler situations, as the hexapod allows the rotational centrepoint of even a single angular optimisation to be placed on the optical axis, at the beam waist, etc. This is vital for reducing parasitic geometric errors, another key to improved overall productivity. Sometimes, very long travel is necessary in one or two axes for loading operations, and in these cases the hexapod can be mounted on a long-travel motorised stage. The hexapod controller accommodates two additional DC-servomotor axes. Alternatively, force-sensor elements can be integrated.

The alignment processes

There are two types of processes: gradient searches, intended to efficiently optimise coupling (and optionally track it to mitigate drift processes, disturbances, etc.), and area scans, intended to localise a peak within a defined region.

Gradient searches

Gradient searches perform a small circular dither motion of one device versus the other, which modulates the coupling. The amount of modulation of the figure-of-merit being optimised (for example, optical power or MTF) is a measure of the local gradient of the coupling. The modulation falls to zero at optimum. 

From the observed modulation, one can mathematically deduce the local gradient via a very simple calculation, such as equation 1. Note that the gradient ∇I falls to zero at optimum.

|ε(θ)|=∇I=(I_min-I_max)/I_min

Equation 1: The observed gradient serves as a measure of alignment error.

Any axes in an FMPA system can perform any of these types of alignments (subject to the physical capabilities of the axes, of course). So, area scans can be performed with motorised-stage axes, which can be very handy for finding first light. Gradient searches are most familiar from transverse optimisation, but they can also be performed, for example, in a single linear axis, which is ideal for localising the beam waist in a lensed coupling, or in a gimbaling fashion to optimise an angular orientation. There are many possibilities. These are highly general-purpose algorithms suitable for all kinds of optimisations, including bulk optic, cavity and pinhole alignments.

In general, FMPA allows different gradient searches to be performed in parallel. Transverse optimisations tend to be the most sensitive and the most affected by other alignments, so transverse routines tend to be relegated to high-speed, high-resolution piezoelectric stages such as PI’s P-616 NanoCube. The high speed and continuous tracking capability of the NanoCube allows transverse optimisation to be maintained during Z and angular optimisations that would ordinarily require the time-consuming, looping sequential approach.

Area scans

Scanning an area to determine the approximate location of the highest coupling peak is useful for a variety of tasks, such as:

• first-light seeking;
• profiling for dimensional characterisation of a coupling, which can be an important process-control step; and
• localising the main mode of a coupling for subsequent optimisation by a gradient search, thus helping to prevent locking-onto a local maximum and is very powerful.

In addition to reducing the area scan to a single command, FMPA controllers have automatic curve-fitting capabilities built in, plus a data recorder that can capture the profile on-the-fly for later retrieval, analysis or databasing. FMPA area scans are very fast, 300 msec or so for typical NanoCube applications and loads. The curve-fitting capability can fit a Gaussian to a fairly sparse scan (meaning an especially fast scan), allowing good localisation of the optimum coupling point without taking a lot of time to do a really fine scan. Another capability is finding the centroid of a flat-top (top-hat) coupling, such as seen when probing a deposited photodetector with a single-mode fibre. This allows the scan to terminate with the fibre at the geometric center of a flat or tilted top-hat coupling.

FMPA’s area scan options include single-frequency sinusoid and spiral scans. These are much faster than traditional raster or serpentine scans since they are truly continuous and avoid the settling requirements of the stop-and-start motions used in traditional scans, and the frequency can be selected to avoid exciting structural resonances. A constant-velocity spiral scan may also be selected, allowing data to be acquired with constant density across the spiral.

Example 1: wafer probing of angle-insensitive devices

Even in the simplest case of a short waveguide device with just one input and one output, the steering interaction mentioned above can present a frustrating process bottleneck. Add the additional alignments necessary for angle-sensitive couplings and array devices, and the situation would grow complex and time-consuming very quickly. Parallelism mitigates all this and makes short work of the task.

For this example, first consider a planar waveguide device with a single input and single output, both accessible for probing via diffractive couplers. Many thousands of such devices are commonly fabricated on large wafers and therefore, throughput is very important. The diffractive couplers typically project the waveguide’s input and output out of the wafer at a typically 10-25° angle from the vertical. Often, lensed probe fibres are used, so there is a fairly distinct optimum separation along the optical axis. Wafer probers of good quality provide wafer placement accuracies much less than the 100 x 100 x 100 μm field of view (FoV) of the NanoCube, so first light seeks are generally not needed on a per-device basis in probing applications.

Note that the optical Z axis is at an angle to the mechanical Z axis for normal mounting of the stage stacks: Z__optical ∦ Z__mechanical. Usually, it is not desirable to tilt the motion assembly to accommodate the angled optical beam, since the mechanical X, Y plane should remain parallel to the wafer to avoid collisions. Consequently, optimisation motions in mechanical Z must be accompanied by compensating motions in X and Y to keep everything aligned.

This is an ideal case for parallelism. From the generic list of alignment routines compiled above, the first four apply:

1. Transverse optimisation routine, input.
2. Transverse optimisation routine, output.
3. Z optimisation routine, input (beam waist seek).
4. Z optimisation routine, output (beam waist seek).

Using traditional, non-parallel alignment technology, the conventional approach would be as follows:

A) To accomodate the Z__optical ∦ Z__mechanical, loop:

1. to accomodate the steering effects, loop:

    a) align one side to maximise throughput;

    b) align the other side to maximise throughput.

11. move in Z and evaluate if the move direction improved coupling.

B) Repeat the above steps until the loops are optimised.

Overall time required is often many tens of seconds.

Using FMPA, the process is much simpler and typically takes only 1 second. Fundamentally, one defines gradient searches 1–4 from the list (again, this only needs to be done once, though any process can be modified or re-defined freely); and for each device, issues the FRS command: FRS 1 2 3 4. Execution is typically complete in a few hundred msec.

A single E-712 controller supports up to four P-616 NanoCubes, which can even be deployed on different workstations; they do not all need to be processing the same device.

Tracking and the completion criterion

A signature advantage of the gradient search is that it cannot only optimise but also track its optimum. If one has several gradient searches operating on a device, all can track simultaneously. Alternatively, one can align and then stop and hold position at the optimum.

This criterion—whether to align-and-stop or keep tracking—is an example of a parameter that can be adjusted in the process definition to fine-tune the process’ behaviour to meet specific application goals; there are many such options. Since the process depends on the instantaneous gradient of the coupling, it is natural to define its stop point in terms of the gradient. This is called the minimum level (ML). Setting the process’ ML parameter to 0 means it should never be satisfied and should track until commanded to stop. This is very useful for accommodating drift processes, such as in elevated-temperature testing. Setting ML to a small but non-zero value causes the gradient search to terminate at the observed optimum position. Note that ML=0 tracking should only be performed with flexure-guided mechanisms due to the potential for mechanical wear in mechanical mechanisms.

Example 2: wafer probing of angle-sensitive or array devices

Building on Example 1, it is possible to accommodate angle-sensitive devices and array devices by using a hexapod rather than an X, Y, Z stack of stages for the long-travel motion. For many applications, a hexapod will provide sufficient resolution and speed, otherwise (or when continuous tracking is needed) a NanoCube can be attached. Again, the modularity of the FMPA architecture yields considerable flexibility.

Alignment hexapods have many advantages over a stack of conventional linear stages and angular positioners such as goniometers. First, hexapods are full six DoF devices, and their rotational centrepoint is freely settable anywhere in space. This means one can rotate about a fibre tip, a beam waist, a waveguide axis or any other optically desirable point in space.

PI hexapods present a sensible Cartesian coordinate system to the user (X, Y, Z, θX, θY, θZ) and allow that coordinate system to be easily cast and rotated as desired. Among other things, this means that a hexapod can be mounted on an angle bracket to minimise overall footprint, while still having an X, Y scan plane parallel to a wafer or other important datum surface.

Again, the FMPA area and gradient search capabilities (for linear and angular axes), data recorder and profile fitting functionality are built into the controllers, yet costs are typically less than a stage stack of equivalent resolution and motion performance.

Consider the case of an on-wafer device with an array input, output or both. On the sides with the arrays or other angle-sensitive elements, the hexapod-based alignment engine is mounted. Using traditional, non-parallel alignment technology, the conventional approach to optimising this would utilise the following even more complex looping sequence:

A) For the first channel, to accomodate the Z_ optical ∦ Z_mechanical, loop:

1. To accomodate the steering effects, loop:

• a) align one side to maximise throughput;
• b) align the other side to maximise throughput

11. Move in Z and evaluate if the move direction improved coupling

1. Repeat the above steps until the loops are optimised.
2. Increment in θZ.. It is unavoidable that the transverse alignment will be degraded by this.
3. Repeat A. Evaluate if the increment improved coupling in the Nth channel.
4. Repeat A-C until both the first and Nth channels are optimised. For most applications, this will mean the entire array is optimised.

Overall time required is typically multiple minutes.

Using FMPA, the process is again much simpler and vastly faster. As in Example 1, one defines gradient searches 1–4 from the list, then one defines either a single-axis sinusoidal (area) scan or a gradient search in θZ; let this be called process 5. (The gradient search requires some initial coupling, but the area scan does not, so the choice of one or the other depends on the application, fixturing, device consistency, etc. Also, setting soft limits to prevent a gradient search from walking away if optical power is cut is both prudent and easy).

Then, for each device, the following steps would be undertaken:

• Set the NanoCube to track continuously (ML=0 for processes 1–4) for the first channel and issue the FRS command: FRS 1 2 3 4.
• With the NanoCube tracking the transverse and Z couplings on both sides of the waveguide, issue the FRS command for the θZ process 5.

Execution is typically complete in less than a second.

Addressing gimbaling (θX/θY) alignment is similarly simple and fast with FMPA.

Summary

This is intended to be an illustrative but not exhaustive description. Similar sequences can be quickly devised for packaging, chip-test and other applications. There are other ways of performing each of the examples presented, and application considerations may recommend different approaches or modifications. There are additional options and parameters that should be considered in an actual application. This overview should provide confidence that the productivity of parallelism is accessible in many applications.

Physik Instrumente

www.pi-usa.us