Hello, I am Laurence Montagner. I'm working for Microchip Technology Rousset. I am in Aerospace & Defense Group. I'm going to show you how we use JMP to develop a predictive single event latchup model.
This project, called SELEST, started four years ago to develop an internal SEL prediction tool. This work was funded by the CNES, the French space agency. Two posters were presented at RADECS conferences in 2019 and 2021. In the first poster, it was a new approach and the feasibility of using an analytical model. That's what we are talking about today and that's how I'm going to present. The second poster, it was with a more accurate model using a neural network approach.
The context, to address the new space market that's been low cost, that means that our circuits are going to be launched in low Earth orbit. We are going to reuse COTS— circuits on the shelves —to make them radiation tolerant to meet space agency specifications. We need to analyze a lot of product to know if we can make them [inaudible 00:01:33] tolerant. W e need a model and a predictive tool to gain the time and money before any experimental tests.
When they are sent in space, those circuits are under radiations. We have several source of radiations: the sun, cosmic rays, and Van Allen belts. The sun and cosmic rays emit electron, protons, and ions. As for Van Allen belts, the inner belts emit proton and the outer one emits electrons. Those particles, when they strike our circuits, causes damages.
We have two family of damages. One, the TID— the total ionizing dose—w e don't talk about it today. We are going to focus on SEE, single event effect. We are going to focus more specifically on single event latchup. This single event latchup leads the component to the destruction. That's why we need absolutely to predict this phenomenon.
The mechanism on the single event latchup is very similar to the electrical latchup, but it is not provoked by the same causes. In the single event latchup, this is when heavy ions is striking sensitive devices as inverters [inaudible 00:03:33] . In the worst case, it is striking in the middle of these devices, and it triggered a parasitic thyristor, composed here of an NPN and PNP bipolars.
When the supply of the circuit is upper, that's the Vhold of this parasitic thyristor. At that time, the thyristor is still on, and we can lead to destruction. As you have understood, this parameter of Vhold is very important for us.
I just talked about energy, the energy of the heavy ions. There is a physical quantity very important, and is criteria of a space agency. This is the LET, linear energy transfer. This is amount of energy lost in the matter by unit track lengths. For ESA, European Space Agency, circuit is said to be immune to latchup when this value is above 60 mega-electronvolts centimeters square per milligram.
Our objective is to be able to predict the V hold and the LET threshold of a circuit, so we need to build a model. We use TCAD Sentaurus to run some simulation to build the model with a DOE. For the DOE, we need to define the input and outputs. As you have guessed output, V hold and LET threshold. Regarding input, for the first try, we decide to define four output, two from the process, epi thickness and epi do se, and two from the design, the length between the two well of the test structure of the inverter, and the length between the top and the well .
Keep in mind that if the Vhold obtained by simulation is upper of the supply of the circuit, at that time, the circuit is immune to SEL. If the Vhold is inferior to the supply of the circuit at that time, it is possible to have a single event latchup, and we are very interested to know the LET threshold.
In the flow, we are going to build a DOE with JMP, a Full Factorial DOE. We are going to input this DOE in the TCAD Sentaurus. We are going to run our simulations. We are going to take the output, Vhold and LET threshold. We put all results and input in JMP. We are going to screen data and build a model with JMP.
Now, let's go to a JMP data table that we use to study the feasibility of using an analytical model. This is the table with our four input and two output here, and the full DOE, one value per color. To input here, we can see a preview of what we have in each distribution. To have a better display and have a better exploration of data, I begin by putting distribution of input and output. We can check that for each input, we have the same number at each value in each input, so it is okay.
We can go at the end, and we can see that for the output, we have something to note quickly, that the V hold is inferior to the supply of the circuit, which is 1.95. I can put the limit on the graph. We can have a single event latchup. We have the value of the LET threshold in here. What we see that some value are upper that 100. We can highlight it to see if we see something special on our other input. We can note quickly that the Body Ties for one value, whatever other input value, are the source of the LET threshold upper of 100. If we can see upper value for the LET threshold upper of 60, there is not the same effect, maybe, for the EpiThick. With this first analysis, we can explore our data and have an idea of what we have.
Second graph to be plotted for this analysis. It's a very interesting plot. Variability/A ttribute Gauge chart. We are going to plot all our input in X, and in Y, our output. Now, we can analyze first if we are main effect or interaction. We can connect cell means, and we can see that there is maybe a problem on the T CAD results.
It is not a problem. We can continue our analysis and study to know if it is feasible or not to have the analytic model because it is just one dot, but we must note what happened here. We saw it quickly with this kind of graph. If there is a problem on our foot, if we have all our data for each condition. We have all data, but this one is to be analyzed.
As for the Vhold, there is no problem on our output. We can see that what we know of the physics, that when the spacing between the two wells is higher, the Vhold is increasing. This result is consistent. If we see at the BodyT ie, the Body Tie is increasing, the Vhold is decreasing. That's what we know, too.
Here we can see what we have already remarked on the previous analysis. Here, for this value of Body Tie, for the LET threshold, we are at a value of 100 mega- electronvolt. That's what we already noted. We have a trend that when the [inaudible 00:12:07] is increasing, the LET threshold is increasing, and it is in agree what we know. It is interesting. We can say that with this graph on Vhold , we have a main effect of the [inaudible 00:12:29] .
We can keep the same graph, same analyzing, but by changing the order of the input. Re call , better. Remove this input, recall, and we can have the graph. What we can see on this graph, this representation, by just changing the order of the input, that the EpiD ose, for some value of other input have no effect for some condition here. So we can deduce there is interaction between inputs. For LET threshold , an interaction slightly here and no effect for other value.
Now, we have checked, we have no problem on our data. We have seen we have some main factors. We can, by curiosity, use another tool. It is a p artition to know which input is the first to appear. Then on the LET threshold, what we see, so Body Tie first, that's what we have seen, and the EpiThick. We can continue by [inaudible 00:14:24] BodyT ie, too. Body Tie is very important. As for Vhold, the BodyT ie is important. A fter the spacing is…
Now, we have a more accurate idea of what we have in our data, we can go and build our model. We put our input. I don't know what I have done. I remove our input. We have run a Full Factorial DOE . We try this model. We put our output, and we can run it. We can run the model. We can see that I missed something here. I already prepared something. Fit Model, EpiDose , Macros, Full Factorial. Okay, run. It's better now. Or is something wrong?
I prefer this presentation with this plot, Actual by Predicted Plot. In this plot, we can see that the model is not really satisfying because we have a lot of dots far from the red area, even if the R-s quared is at 0.81, it's not fully satisfying.
For the Vhold, we have the same remark. We can have a look at the Profiler. The Profiler can show us the good trend, but we are not going to have a long time on this. We are going to look at another model to know if we can have an accurate model than this one. We are going to try another one. Not so far. We are going to take our input, and we'll try Response Surface, and our output, and we run it. We are going to remove EpiDose because there is no effect.
We can see that this model is better, the R-s quare is better. The V hold is a good model. Now, we have a look at the Profiler. The Profiler, the trends are good for all parameter. We note that there is maybe something to do because we see in this part of the curve, the LET threshold value are upper, and it is not in agree with the physics and what we know, of course. As I have already said, it is a first try. We need to work and work as we've done on the inputs and accuracy of the LET threshold to have better results here. You can see now in the EpiDose, we have quite no variation, a little variation here.
However, this kind of graph is very interesting for us because when we want to make a radiation tolerant circuit, we can use it to help us by given a set of value. We can use here the Desirability Function. Now, we are going to set our desirabilities. Here, we want to be upper than 60 as a criteria of ESA. I put 60, 80, and 100, match target. Here, for the Vhold, we want to maximize it. I put this value, 2, 2.05. Now, I'm going to maximize desirability . Of course, it works. And n ow, not. Match target. Target . Now, maximize this desirability. Now, we have a set of value we wish because here, we have a range of value in the Body Ties here. We can play with it. Here, we can play here with the spacing.
The other representation and the tool we like using in JMP is the Contour Profiler with some fixed parameter. We can have a contour plot here. If you want to change a design value, I put in axis the Body Tie and the Spacing _AC. We are going to put our famous value of 60 for the LET, the Lo L imit at 60. To have a LET upper than 60, we know we can have a value of Body Tie up to, if we take the cross here, about 6.5. We can use the range of the spacing.
This tool, this representation, is very interesting for us to make our product radiation tolerant. This is not the last model we implement, but I can show you that in the red triangle, we can save the prediction formula in the table so that after, we can take it and encapsulate it if we want.
Okay, that's all for me on JMP. Let's go back to the presentation. With this method, we built another model, this one using a neural network. The prediction obtained with this model was compared to a experiment on circuit. We can see that when the experiment shows there is no single event latchup, the prediction by the SELEST, the internal tool, said the same thing, there is no latchup.
When experimentally there is a latchup, SELEST, even if there is a difference between experimental and prediction, show there is a latchup. For us, it was a good result. Not so accurate. That's why the work continue on this model to have an accurate model. That's why we continue working on DOE. We do it per technology node for having a better accuracy of this model. Okay, thank you for your attention.