Optical Monitoring Drives Yield

The semiconductor industry relies on an arsenal of well developed fabrication techniques to realise an ever increasing variety of electronic and optoelectronic products. Two of the main techniques employed are vacuum precision coating for the deposition of multiple layers of material on a substrate surface, and dry etching, also known as plasma etching, for the controlled selective area removal of material to define circuit elements.

In recent years, the overwhelming trend has been towards higher performance products and higher process yields through ever tighter process control. For example, crystal monitoring techniques used in the production of optical coatings may achieve an accuracy figure of a few percent. This has proved adequate for a number of multilayer products but fails to achieve the desired precision and repeatability for a wide range of precision optical coatings such as WDM filters, laser facet coatings or gain flattening filters. In these cases, optical monitoring is rapidly becoming the technique of choice.

Optical monitoring generally involves illuminating the sample being processed with either a laser or a white light source and monitoring the interference pattern from the growing, or etching, surface. At Intellevation we believe that a laser based system offers the most robust and flexible solution for dry etching applications. Unlike other techniques, such as optical emission spectroscopy or SIMS, the laser provides a high intensity (i.e. low noise) ‘probe' that is independent of the plasma chemistry, the gas flow rates, the chamber geometry, the sample size and the amount of unmasked area.

For instance, many MEMS etch processes employ high density plasmas to realise high etch rates. An example of this is the ‘Bosch' ICP process used to achieve deep vertical etching of silicon. In this process, the gas flow rates and the plasma density are relatively high and therefore it can be extremely difficult to observe the relatively low density of etch products using OES or SIMS. However, laser reflectometry achieves extremely clean signals with good modulation depths.

Unfortunately getting a good signal alone simply isn't enough to seriously impact upon the yield of a dry etch process. The real goal is to understand and automatically ‘follow' the signal in order to achieve the correct endpoint in an intelligent, robust and repeatable manner. And this is where Intellevation comes in.

Shape Recognition for Dry Etch Monitoring
When etching through a bulk layer, it can be shown that the reflected signal varies in a readily predictable manner with turning points occurring every _/4n, where _ is the laser wavelength and n is the refractive index of the material being etched. However, it has also been shown that the signal from anything but the simplest of wafer stacks, such as those required for III-V optoelectronic devices, varies in a far more complex manner with turning points arriving at uneven spacing and with amplitudes varying in a non-obvious way.

For this reason, the company has put as much priority into developing unique software analysis algorithms as it has developing the monitoring hardware, pioneering the integration of predictive modelling into its laser endpoint products. Before the etch run is commenced, this facility, in its most general form, accepts the wafer layer materials and layer thicknesses as inputs and from them, predicts the complexities of shape in the reflected signal as a function of etch depth. Furthermore, Intellevation has integrated patented shape recognition algorithms that enable the endpoint system to automatically follow the live signal in terms of the predicted response. Thus, at any point in the process, the actual etch depth is known by reference to the model, and the endpoint can be called with extreme accuracy.

The above technique has found great success in the etching of complex wafer stacks such as those used in the fabrication of GaAs lasers, InP LEDs and even GaN devices to name but a few. Generally etch depth accuracies of better than ± 5nm in a 1,500nm etch can be routinely achieved using this technique.

In addition, the shape recognition capability has led to some unexpected applications. Such an example is the selective area patterning of SiO2 on top of Si. In this case, the requirement is to rapidly etch the SiO2, and to endpoint the etch process as soon as the Si surface is exposed. A rapid etch process means that a high etch rate selectivity between the SiO2 and the Si cannot necessarily be achieved, and therefore to avoid removal of any exposed silicon, or even damage to the newly exposed silicon surface, the etch must be stopped with extreme precision and speed. A further complication is the fact that the precise starting thickness of the capping SiO2 layer is often variable from wafer to wafer and the cost of individually measuring this is prohibitive.

In fact these algorithms can be used to detect even more subtle effects such as monitoring the etch through various SiO2 layers having nothing more than different dopants and/or doping densities. The minute changes in refractive index and/or etch rate result in subtle changes in the shape of the reflected signal as a function of time. These changes, although unobservable to the naked eye during the etch run, are readily picked up by the shape recognition algorithms, again resulting in a positive endpoint.

Process Simulation for Optical Coating
The big differences between monitoring for dry etching and optical coating are the dimensions being monitored and the number of endpoints being detected. Dry etch applications typically have one endpoint at a depth which is usually much greater than one wavelength. Therefore an endpoint accuracy of a few nanometres in an etch depth of 1 micron results in a very small percentage error.

However, the growth of a film stack requires an endpoint for each and every layer, where each layer may typically be around 70nm in thickness. Depending upon the nature of the film stack, the effect of any ‘cutpoint' errors can rapidly become compounded with each subsequent layer, and consequently, the product can soon become unmanufacturable as the number of required layers increases. This is the main reason why optical monitoring is crucial to achieving high performance optical coatings.

The optical monitoring hardware uses a white light source to illuminate the surface of a growing film stack. A detector module contains a monochromator enabling the optimum wavelength to be used to monitor each layer. The system automatically changes wavelength, changes test glass, controls shutters and sources, all whilst monitoring the interference signal from the growing film stack.

Standard ‘stand alone' thin film design packages model the optical signal from a growing film stack, thereby allowing the coating engineer to determine when to make the ‘cut' at the end of each film. The output from these models is calculated from the optical and physical properties of the materials such as refractive index and film thickness.

However, in reality the optical signal from any growing film will also be modified by the specific foibles of the coating tool and coating scheme being used, as well as by the hardware and software associated with the optical monitoring system. Some of these effects inevitably lead to parameter fluctuations not accounted for by the ‘ideal' model. For example, in a typical e-beam system RF noise spikes are often generated from the switching of high energy supplies.

Also, film deposition properties, such as the instantaneous deposition rate, film composition, uniformity and ultimately the refractive index, can also vary depending upon the stability of the e-guns. These run-to-run variations can have a dramatic impact on production yield and are not picked up by the ‘ideal' models.

The FilmSimulator package reads in the preprogrammed film stack design along with its associated process scheme details including monitoring wavelength changes, test glass sequencing, filtering schemes, cut-point detection algorithms, etc. The user can then program a number of coating tool dependant parameters specific to their system.

The simulation package simulates the coating run, calculating the cut-point errors on a layer-bylayer basis, propagating them through the stack as it is built and displaying them in an easy-to-use graphical form. This gives the coating engineer a powerful tool to see inside the process and determine exactly where deviations from the ideal are likely to occur.

The simulator can overlay the output from many simulated runs giving valuable information on the run-to-run variability and therefore the process yield. Rapid re-optimisation of the design can then be undertaken off-line and the improvements instantly observed thereby eliminating many time consuming test runs on the actual coating system. It should come as no surprise that there are manyfold error sources to confound the transition from design into reality. What is surprising is just how difficult it is to guess the impact of each of these sources on the final outcome.

Conventional filter design packages are particularly poor in this regard, and FilmSimulator stands alone as the only software tool with such robust capability.

The interplay between the accuracy of the cut achieved for an individual layer, and the optical performance of the entire film stack can be extremely complex and sometimes downright counter intuitive. There are often many possible variations on a particular design, but not all are equally manufacturable.

We come across awkward designs where the chances of hitting an acceptable process after even the 10th coating iteration are slim. Using these process simulation tools we've been able to rapidly pinpoint problem areas within the customer's coating scheme and identify alternatives with quite dramatic impact on performance and yield. We have found that an hour spent on the process simulator can save the customer many days of process development in the lab.

So optical monitoring has a bright future when it comes to dry etching and precision optical coatings.