0:13 And good morning. I am my name is Greg. Can Can you hear me okay? Yes, I can. I am primarily a mechanical engineer and I’m happy to be here to present today a probabilistic design toolkit we have developed at Medichec to manage risks in product development. Okay, so the most common request we get from our clients is, “Can you help us improve our products? Can you make them stronger, lighter, cheaper, better?” And this is what we do at Medichec.

0:42 We use different in different ways. We use We develop digital twins. We use computational modeling. We can perform physical testing. We have advanced design capabilities. But essentially, in order for us to do our job well, we need first to dive deeply into how a product or a design performs before we tweak it and we optimize it. And this is the core service we offer at Medichec. Now, throughout the years, we have realized that in many cases, by adopting traditional design and design assessment methodologies, our approach that are in part are limited.

1:21 So, my aim today here is to challenge these traditional design and design assessment methodologies and convince you that by embracing a bit of uncertainty into the design process, you have a better chance to understand how your product will perform and therefore optimize it more efficiently. Let’s start with how this is done traditionally. How do we go through design operation? How do assess designs the traditional way? So, usually, we would set specific values of input or input parameters that represent in-use or abuse cases and would perform either digital or physical testing.

2:00 And we would exaggerate the values a bit to account for any uncertainty. Then we would get the result from this testing and we would compare against again specific values of said key performance algorithms. Putting them as red lines we cannot cross. The again, we would compensate a bit of we would exaggerate the values a bit to account for any for any uncertainty. Now, that approach is deterministic and it assumes that the design will either fail all the times or it will never fail, which is far from what’s happening in reality.

2:38 But values of variables like one factor tolerances, material properties, even the success rate of quality control measures fluctuate and they can fluctuate a lot. So, is there a better way to incorporate the that fluctuation into our design process rather than just exaggerating our testing conditions? This is why at Mettler Toledo we have developed the probabilistic design assessment toolkit, where we can incorporate uncertainty represented by a statistical distribution into specific input parameters of a computational model to perform finite element analysis.

3:18 Now, we can then go directly and develop a lot of models of different combination of values of the parameters, taking into account the statistical variability. Following your model parallel approach, or we can do a more rigorous step beforehand, which is which is to evaluate the effect to evaluate the effect of each of the parameters and these uncertainties and maybe eliminate those that don’t matter and matter and focus on the more critical ones by following a design requirement method like Taguchi.

3:51 Both ways will end up in in a stack of simulation that will be performed that will be then realistic product trials essentially. Now, the output of this tool kit is not very different to the usual output you get with computational modeling. So, we still get contour plots of stresses, we can get forces, strain deflection, but now all these come with a statistical output. So, we know how likely is to reach this value, we know what the average result might be, and we also can plot results on specific confidence intervals.

4:26 If we correlate all those together, then we can also get an understanding of how likely it is a product or a component or a design will fail. As you can see here, this contour plot represent the likelihood of failure to occur at a specific location of an alloy wheel after driving for 100,000 mi. Okay, let me go through a case study and hopefully explain the process in more detail.

4:52 So, this is a titanium bracket that is used to mount sensors on a car device. It it consists of a main body where some board and camera are mounted on, and the two flaps where the sensors are mounted on. Now, this is additively manufactured following a specific process, and the main objective of redesigning was to make it as lightweight as possible while achieving key performance targets. And one of the most important ones was to make sure that this component can survive a drop.

5:29 And and it won’t break or or fail in any way. Now, the usual way we would go about this and we would use computational modeling to assess that would be to assume a worst-case scenario. Assume a worst-case drop height. If someone is tall enough, maybe they can reach a height of 2 m. Assume that they’re going to land on the stickiest floor type possible, which probably is going slab, and let’s test a maximum three orientations, which usually are the three polar orientations defined by the assembly data of the product, which can be quite random.

6:06 But in reality, we know that it’s this height is very unlikely. We got the data from the sensor on the device showing that it’s very unlikely to reach any height greater than 1.2 m while being used. This is This is intended for indoor use, so it is more likely to have a different floor type in your living room than a concrete slab. And finally, any drop orientation possible, which can be dropped anyhow.

6:32 So, that’s why we incorporated We incorporated these three drop loading input parameters into our probabilistic toolkit, and we associated our certainty to them based on data we got from sensor literature or previous failures. We also added an uncertainty to the material properties of titanium, and we all know how variable it can be, especially in additive manufacturing. And we use this finite model to perform 1,000 drops, 1,000 different simulations of different combinations of these parameters representing real trial real drop drop trials.

7:10 Now, as I said earlier, accuracy in order to optimize we need to take on 10% of something that works. So, this is a threshold under load of the component being dropped, and what’s happening is that we have a global motion, the whole thing is being dropped, and first the first thing that’s moving is the main body, and then the flaps follow with a delay due to inertia effects.

7:32 This is called This is causing some stress to rise at the transition between the main body and the flaps. And the way we can reduce the stresses is by stiffening up that region. If we make it stiffer, if we make it as rigid as possible, then we can alleviate the stresses. Now, another key performance target for this component is to make sure that it retains the same relative angle between the flaps during in-use conditions.

7:55 And this is very difficult, especially when there’s thermal loading, when this is heating up exactly because there are a lot of layers stacked up above and below the main body of the component, and they won’t be expanding in a different way than when they’re heated up. This thing wants to bend. So, the main body tries to bend, and that is causing the flap to be pulled in, basically.

8:18 So, where the way to eliminate that and make sure that that kind of deflection of the main body doesn’t affect the alignment of the flaps is to isolate the flap from the main body. So, we need to make that transition more flexible. And And the reason I’m showing this is this highlight how complex the optimization process can be. We have two conflicting requirements. How How do we What do we do about it?

8:41 Do we make it stiffer? Do we make it more flexible? Essentially, probably we need to find a sweet spot between those two. Okay. Let’s move on now to the assessment of the original design using the traditional analysis approach and our probabilistic design tool. The traditional way we get the pre-orientation that we tested, and early on we realized that when this component is being dropped at its part of the machine, we have big stresses on that zone here.

9:11 This means it’s going to fail. We need to mitigate that. A redesign is necessary. And this is the main method we get we get with the traditional analysis. Now, the first piece of information we get from the probabilistic toolkit is this kind of heat map. So, here each dot represents one simulation that was performed. It is arranged So, it is arranged based on the direction of the drop.

9:37 That’s why you have the three arrows here and then go like that. And the color of each dot represented the peak stresses that were recorded in the model for these simulations. So, quite early on, if we look at the from the top as you see here, we realize that the darkest red dot, which means that the highest stresses in the simulation that showed the high stresses, it’s not aligned with any of the principal directions we tested traditionally.

10:08 So, although the traditional methodology was very conservative and and we show we we show that, it still manages to miss the direction that is actually causing the worst-case stresses to the model. The other thing we can learn from this toolkit, if we look at the finger rate, it’s very low, it’s less than 0.7 percent, and it’s very localized. So, it is not something we should really worry about.

10:35 This is very easy to improve and it can also be quite possible it is not a problem anyway. All right, so and this is again conflicting. The data approach tells us that this is fine, if something it might be over designed. This traditional approach is telling us that this is problematic. We need we need we need to redesign anyway. Okay, so we use the information from the traditional analysis and the probabilistic analysis and we did optimize the design.

11:02 And the first optimized design we used, I’m going to talk with how we used the traditional analysis information to optimize the design. This is what we ended up with. We achieved a 49% volume reduction, and the way that we did that is by implementing two supporting structures, one spine, vertical spine, that controlled how we think pivots across across the component, and one horizontal rib group that connect that spine to the blossom.

11:35 So, if you remember earlier, I was saying we we we we need to find the sweet spot between how flexible or stiff that is, and by tweaking the the thickness and the length of basically this structure, this support structure, we managed to find out that we can see that very easily on the drop result. So, this is the original drop result, and you can see that these fixed stages are eliminated.

11:58 And and and and that will that will be the end of that task if we’re following the conventional design approach. Now, this is the design the optimized design we ended up with following the probabilistic using the probabilistic information we got from the probabilistic analysis. We achieved a 65% volume reduction with a 60% more. And and the way this works is that exactly because the probabilistic toolkit showed us that the the the transition from the main body to the blossom is not as heavily loaded as traditional methods suggest.

12:35 We had the opportunity to explore a very different design concept. We had the opportunity to make that very flexible. And and adopt a new design. So, the bottom deck remained the main body where all the electronics are mounted on. The top deck is where the two sensors are mounted on, and the top deck is is stiff is stiffened up to make sure that we never lose our kind of alignment between the two sensors.

12:58 And we were able to make the the transition very flexible, but it represented a suspension system, basically. So, when it’s being dropped, it’s back, we can absorb the shock. Now, this is not the result of an automated process. This is not the result from an automated software. This is pure AI to human diligence. I think it would have been very difficult to get that almost impossible to get that through an automated software exactly because we cut it with the bracket, and we ended up with a suspension system as this problem we needed.

13:27 So this is us understanding well how the product performs. How we design feature affects, how it going to be loaded, and finding the right solution by manually using nerve surfacing tools to to get to get the right to get the right response to the problem. Let’s see how this perform. So if we look at the heat map again, same heat map as before, this is the original one, this is the new one.

13:56 You can see that the dark red dots are now more in the orange zone, which is great. But you can also see that in all other directions, in most of the cases we see an increase in stresses. Still still still we are comfortably under the limit of the material, so all good. But that’s that’s shifting the design, also shifting how the product performs. So we have a more stressed response.

14:22 We have a better optimized solution for multiple loading directions rather than just one rather than the worst case one. So basically in in a few words, each gram on this design matters more because more of the volume is being loaded in every direction pretty much. And the failure rate was also very low and very localized as before, less than 0.4% and again very localized that can be very easily addressed.

14:51 Now for comparison purposes, we also tuned for the traditionally optimized design through our probabilistic toolkit. And this is the result on heat map that you see here on the left and we’re going to compare to what I just explained about how the shift of the design helped with that. So so what we see here compared to the original is that the dark red dots are now again in the orange region, but also everything else seems to have been scaled down.

15:18 So whatever here is in orange is more in the yellow. Whatever here is green is more on the blue. So, we see a scale down effect of of basically we optimized for these worst case loading conditions and we managed to scale down stresses throughout so through for any other simulation. So, yeah, which is different than shifting the performance. So, this is a shift in the performance, but it optimized for a drop while this is a scale down to make sure we’re under the limit.

15:56 And and both of them ended up with a very similar failure rate control block. Very low, very localized. So, so, so good. Okay, so as I explained from the beginning for us to optimize is to understand. So, we have developed a tool that that could give us a better quality information. Give to the engineering and the design team better quality information about holistic performance tutorial. Which allowed us allowed us to find the leeway required we get up 60% but more importantly see the design console to something that we could optimize for how this load would act on it.

16:34 I’ll be very happy with the talking points. So, please come and find me during the break over a coffee or even better a beer. The yeah, with any comment or questions. Thank you very much for your time. To learn more about the CDFAM Computational Design Symposium, access the archive of previous presentations, interviews with speakers, and information about future events around the world, visit CDFAM.com.