Automated Process Control of Ion Implantation Using Linear Regression (2021-EU-PO-753)
Victor Kessler, Group Leader Process Engineering, First Sensor
First Sensor produces very sensitive optical sensors based on four-inch and six-inch silicon wafers. Especially sensivitve are our Avalanche Photo Diodes (APD). These APDs have a critical breakdown voltage which is adjusted using a special external high-energy ion implantation process, which I control. The breakdown voltage is extremely sensitive to small dosage variations of the implantation process and needs to be controlled with data from our wafer prober measurments. This data is acquired at the end of the manufacturing process; the time lag between implantation and measurement is usually about two to four months! In order to optimize and control the implantation process, I set up the following JMP script/routine:
1. Script goes daily through all new wafer prober measurements.
2. Wafer maps are generated and automatically saved for all products (useful eye candy).
3. Wafer prober statistics are saved (median breakdown voltage).
4. Statistics are joined with implantation parameters.
5. Outliers are excluded.
6. Linear regression is performed.
This automatically yields the needed dosage for a specific breakdown voltage.
Auto-generated transcript...
Speaker | Transcript |
| Victor Kessler | Yes, hello, my name is Victor Kessler. I'm working for the company First Sensor, which just has been acquired by TE Connectivity. |
| And I'm presenting today about the special high energy ion implementation process and the process control about it using linear regression with JMP. | |
| So First Sensor produces silicon based pressure and optical sensors and our well known special products for detecting light avalanche | |
| photodiodes. They are used, for instance, nowadays, and in lighter systems to detect ranges and one of the future applications might be autonomous driving. | |
| And there, it is necessary to detect small amounts of lights and thus quite fast, and therefore our avalanche photodiodes | |
| can be used. And one of the special characteristics of those diodes is their breakdown voltage. That's the | |
| special characteristics for our customers, where they need to adjust this breakdown voltage in order for the application. | |
| And so we need to adjust that and set this breakdown voltage just right within specification, obviously, and this is quite a sensitive process for those diodes. | |
| We use this...we control this breakdown voltage with the doping profile of those photodiodes and this, again, is controlled with this high energy ion implementation process. So what is happening is that, obviously, we are fabricating those | |
| photodiodes with wafer processes and microtechnology, so they have a long route through all of the processes. We do add in one step, we | |
| reach out to our partners at Helmholtz-Zentrum Dresden-Rossendorf, which is located in Saxony, within Germany, and they have a special facility in order to achieve a really precise control of dosage at high energies. And this ion implementation process is done at a quite a high energy. | |
| I will not say the exact | |
| energy right now, but it's something you cannot do with usual ion implanters and the dosage is really low when it needs to be implanted with wafers, which is another challenge. | |
| And it is so sensitive that there is no real in-process control available at the Helmholtz-Zentrum Dresden-Rossendorf in order to check | |
| the exact dosage. So what needs to be done is to establish a relationship between different dosage levels with has been implanted and then our actual results in the end of this breakdown voltage. | |
| And this is | |
| also depending on the wafer size, since we are fabricating wafers into different sizes. | |
| We have two different sizes and we have different product areas where the photodiodes | |
| fabricated for customers for detecting light in different wavelengths. And in the end, we end up with like 10 necessary process controls, | |
| which needs to be established. And another | |
| challenge of this process control is that the process itself, the ion implementation and a measurement of the breakdown voltage, has a timeline of many weeks, sometimes months and so this adds up. | |
| Yeah, it's a challenge. And what I came up to with in order to be able to have better process control and more precise control is using JMP | |
| in order to automatically control this whole process. So what was establishing | |
| most...well in...with many scripts was a routine where everyday | |
| to task scheduling a script automatically at nights, starts to search for the latest way for proper measurements, which are being done at the backend of the process. | |
| And it collects all new wafer prober measurement results. This is necessary, since we currently do not have a database for this, so I need | |
| a script searching for those results. And it collects the data, it joins it with the implantation parameters, which are saved in another table, | |
| and filters outliers and saves statistical values, especially the medium of the breakdown voltage of a wafer to quite comprehensive JMP data table. | |
| Also on the way, since I thought, well if the data is there available, I rather also generate some wafer maps for visual inspection and automatically save them as well | |
| to later inspection. And this can be seen here on the left side, and I will now enlarge the image. | |
| So yeah. | |
| Those are the...on the left side is the breakdown voltage and on the other side, another characteristic is plotted, which is useful for us for an inspection. And you can already see here that there's quite the distribution of the breakdown voltage up about... | |
| over a wafer. So we have quite some variance of this characteristics on a single wafer and then we have a wafer-to-wafer variance and around to run(?) variance of this breakdown voltage. | |
| And yeah you can see here I not only generated wafer maps, I included the distribution and calculated also teh yield and so on. And some other characteristics I put here in the title, | |
| as well as logos and also metainformation. I made some of it black since it's confidential. | |
| But yeah so we save from every wafer problem measurement automatically, these wafer maps for later inspection and to get a feel for process. Also I can see here the influence of other implementation parameters, so we have quite the... | |
| yeah | |
| ...has a strong feedback loop with Helmholtz-Zentrum Dresden-Rossendorf, and so we adjust the process together with them. | |
| And yeah so that's an eye candy which is produced every night automatically. That's nice, but in order to now be able to better control the process, as I said, we need to really build up a relationship about dosage and the breakdown voltage. And this is done here, where you can see | |
| the relationship between the median of the breakdown voltages of many wafers, with respect to the dose...usage dosage for implementation. And, as you can see here, there's quite some huge variance and not not only with | |
| the single dosage, which is then wafer-to-wafer or run-to-run variation. | |
| And that that makes it quite a challenge in order to produce, but from a statistical point of view, of course, you can become better and better if you | |
| aim for the most the...the median you would like to achieve and get the highest yield. So automatically | |
| the complete database for this is achieved by a routine, which runs every night and, later on, we, or I use the prediction profiler in order to aim for the breakdown voltage the customer wants for their purpose and within their specification. | |
| And yeah so that is the state of today. | |
| It's a linear relationship, which is, in the end, established on a second order. We actually can also include the wafer relativity, | |
| but for now, that's not necessary. It also requires much more runs in order to establish that relation. It's...it was already impossible to improve the yield by two figures | |
| with this regression process. | |
| And as a further, yeah, outlook, let's say, for the process control, I will now also script the | |
| fitting process and export the coefficients to an Excel table again where, in the end, then from an Excel menu, which can also be controlled by production planning. | |
| The | |
| colleagues from production planning can calculate the necessary dosage by themselves with the latest | |
| coefficients in order to get the high yield with respect to the breakdown voltage. Yes, so that, yeah, most of the work, obviously, is within scripts, | |
| but it was a nice journey scripting that. And I want to thank, of course, First Sensor and especially help us to | |
| be able to present here and, as well as Frank Kudella, who was the owner of the process for most of the time and ???, and especially Roman Bottger from the Helmholtz-Zentrum Dresden-Rossendorf. Thanks a lot. |
