0:16 All right. Welcome everybody. My name is Max. I’m from Ansys tech lead for simulation topics. And I would like to talk to you about the new native conjugate heat transfer today in our CFD environment in Ansys. And apply that to optimization for turbine blade thermal performance. I’m joined by Marcus Lemke from Siemens Energy who will do the second part of the presentation and showcase all the technology that we have built in the recent month.

0:54 Let’s scope the engineering problem that they go with pretty much and that is in turbine blade design higher temperatures means higher efficiency of the turbine blade. However, we are already in a stage where the operating temperatures on the blade surface pretty much above what the metal can basically do in its solid state so it’s almost melting during operation. But we need for this to to still work is cooling passages inside of the turbine blade.

1:30 Marcus will explain that in much more detail. So we have this design problem where we want highest temperatures possible for the operation and yet we need to maintain the the metal in a solid state. And this is basically what we want to solve and as a computational design setup or in a setup. Why do you traditional workflows struggle with scale and this is pretty much boiling down to meshing.

2:06 If we have complex geometry, we need to create a surface mesh, create a volume mesh, and get it through the solver. And if a single triangle or single child cell in mesh is pretty much ill-formed, the final volume solver will probably fail or mesh control is not given, the residuals don’t converge, and in the worst case, yeah, everything collapses. So, two other hurdles that come with this, and I hope we can agree on them, is yeah, massive amount of solver time, and also HPC licenses if you want to run a solver on thousands of GPUs or CPUs.

2:54 HPC licenses scale, and you won’t necessarily want to spend this money early on in the design exploration phase. At nTopology, we think we can solve this due to implicit modeling. And this is because, implicit modeling, as we’ve heard before, is incredibly robust. It doesn’t fail. We we have 100% geometry output, and we can couple with physics. And this is where pretty much the the beauty of this method comes into play.

3:30 So, all this robustness and parametric design space that we can create with implicit modeling is not worth it in its own, but is rather coupled to the physical performance of these parts. And this is what we wanted to achieve in in one closed loop. Since the acquisition of LBM, a lattice Boltzmann based solver programmed or yeah, developed specifically for high performance GPUs and provides Ant of Fluids, which is a simulation engine fast enough for optimization in the loops.

4:18 It’s based on simple principles and especially it doesn’t have conformal meshing. So, it’s a perfect play out with implicit geometry. We don’t need no surface mesh, we don’t need no volume mesh. Nor is it meshless. I would rather say it’s a voxel based approach. So, these voxels are yeah, very easy and robust to create. It has rapid solution or solving. This is due to the lattice Boltzmann algorithm being very very well fit for GPU compute.

4:49 It’s a trivial algorithm for the GPU. All the voxels are independent. They updated of its neighborhood and that’s a perfect single instruction multiple data algorithm. It’s seamlessly integrated into Ant of which basically means it’s just a block and it operates on the input, the given implicit geometry and it outputs fields that you can use to post process creating streamlines, calculating average values or even inform your next iteration of the geometry.

5:31 As the geometry is a field and the physics simulation is a field as well. And lastly, we’ve heard from Thomas if you can afford it, you want to do LES simulations. So, also a good fit with the lattice Boltzmann method, the method is dependence in its nature. So, large eddy simulations are good fit here. Well, almost get them for free in the method. What we are going to do now is that we have conjugate heat transfer added to this capability.

6:05 And this has taken quite a bit of research and development for us because one thing particular is his complicated about conjugate heat transfer is the coupling between thermal transport between domains. And the implicit modeling these domains have interfaces, but these interfaces are given implicitly. So, we don’t know where the surface is. We don’t have one way to use a mesh now in order to implement the surface into the solver.

6:38 So, we have to come up with a method to describe heat transfer even with touching fluid solid maybe solid solid interfaces without explicitly knowing where the exact position of the interface is. This is where the total enthalpy approach comes into play. It’s a continuous description of the enthalpy conservation law discretized in the lattice Boltzmann scheme. So, we can basically simulate conjugate heat transfer without giving interface conditions or even knowing the exact position of the interface.

7:17 So, it has automatic flux calculation. And you don’t need surface parameters to be given to the solver. It will figure it out automatically. With this, we have a rapid simulation-driven design iteration, which is basically you get immediate insights, informed You can change the parameters inside of an arc. The simulation fields will pop up and can decide on maybe other recirculation zones. I have a pressure blockage of some of these channels.

7:56 So I can inform the next generation, the next experiment I want to do and you don’t change the software. It’s all inside of nTop. And you have one continuous loop to design your your engineering problem. In this case here, a heat exchanger. Make sure to understand this is currently in better access. We’re very happy that this is included now. We know it’s not perfect. I think a lot of models or all models are wrong, some are useful.

8:28 So we want to improve on this. If you experience any unexpected results from this, let us know. We’re here to improve the solver together for all the use cases. And then there’s one last step since all this is just blocks in nTop and you can wrap up blocks into blocks into blocks, why not wrap all this up into an optimization orchestration loop. So everything that I explained is one workflow and they continue we can continue or reproduce this workflow for a series of parameters that we change.

9:09 We can do this manually, but also what we can do is we can use an optimization algorithm to basically choose parameters for us. And yeah, this is how this setup might look like for you. You can set up an optimization problem. You can have a block that executes the geometry creation, the CFD simulation, the post processing and then calculate for example, mean temperature or maximum allowed pressure.

9:40 And then you have independent parameters throwing into the inner loop. And then the algorithm is in this case running for 50 iterations using a global optimizer, something similar to the use of simulation to find a globally optimum result given to a constraints. And with that, I would like to hand it over to Marcus who give a little view on how this is used in turbine blade design.

10:13 All right. Thank you. Okay, I’ll also use the keyboard. Yeah, so why does this matter for industrial turbo machinery? Let me briefly just summarize actually what what Marcus have already mentioned, right? Like so, whenever we design a new gas turbine, the typical requirements are we need to higher turbine inlet temperature and we need less cooling air because less cooling air means that we can push more mass flow through the combustion path and get a higher power output.

10:51 And this gets risen and risen and risen and risen, going ever higher and therefore it becomes increasingly important to really consider all the trade-offs because usually structural integrity, manufacturability, thermal performance, these are all requirements pulling in different directions, right? So, we really need to apply multi-disciplinary design optimization. What then very often happens is when we want to explore different geometry variations, we exactly have the the problem that was mentioned before also that the B-rep they geometry just breaks.

11:36 Or even if that works, then very often, the getting a body-fitted surface mesh is then the the next problem. And there, we also believe that implicit modeling really has this distinct advantage of really allowing us to explore many more feasible design options because the parameterized geometry is so much more robust. That means with implicit modeling, and now with Entop Fluids, also with the Lattice Boltzmann solver, we have the potential to set up a multi-disciplinary design optimization that is much more robust and also very fast.

12:21 And we can which is additional also very nice to have it all in in a single software without having to also deal with the interfaces between software. So, on the right-hand side, because my company obviously is very shy to to allow us to show actual geometries, and definitely not the the latest and greatest, I was pulled something from from academia just for everyone that is not so familiar with what turbine blades and vanes kind of look like.

12:55 That’s specifically a blade here. So, the air flow comes from the from the bottom. And then goes through these serpentine channels, it gets usually ejected to the tip, towards the leading edge, and towards the trailing edge. And for our first proof of concept, I specifically looked at the trailing edge region. So, I pretty much made a a small model of just what a a trailing edge with a pin a pin fin bank might look like.

That is then the model that I created. Again, you can see the the air is coming in from the bottom where it says velocity inlet and then we have outlets, one through the tip, and the rest of the cooling air will exit through the trailing edge holes which are racetrack shaped in in this case. That’s for the fluid boundary conditions as Max mentioned. Now we’re running a CHT, which means I also need thermal boundary conditions.

14:07 That means I number one, I provide an inlet temperature for the cooling air, but much more importantly, I specify a convective boundary condition on the airfoil surface. And again, I could use simulation data to do that, but I just created a generic example to sh- to show you here. Which yeah, is again the blue the beauty of the of the implicit modeling, right? I have a a field that I create that I can directly map onto the surface as the boundary condition.

And I tried to to vary it a little bit already to show how that could work. So, very often in the center of the airfoil, we have a little higher temperature. I also increased it to the back. That is because the thermal barrier coating there is usually a little smaller and I’m not explicitly modeling that. And just also to show you that you can have a an additional variation in the heat transfer coefficient.

15:04 The heat transfer coefficient also increases towards the end of the airfoil. Then the key of this of this geometry is I really want to look at the pin fin bank. And usually you’re in this in literature all the pin fin banks are very regular and that’s also can be very easily such a geometry can very easily be created as a lattice structure with just a a single column.

15:37 And usually there’s a block for that in in in Ansys and usually you would always start probably with at least a regular cell map doesn’t need to be equidistant in all the all directions, but definitely I will be regular. So it will the cell map will always have equal spacing in in each direction, right? But what we see very often with such like when when these geometries were developed and experimentally investigated, it’s always assumed that the approach flow well basically comes from from from the left or here at the top and and flows to towards the the trailing edge, which is not the case, right?

16:23 I just showed you that the air comes in at the bottom and then it turns 90° and goes out. And that already tells you that the the flow will behave in reality much differently than what you had in the experiment. So the idea here is okay, can I actually warp the cell map in a way that I find a better distribution of the pins so that I will get lower metal temperatures.

16:51 And in order to manipulate and warp the cell map, I’m using the Fourier Fourier series that then creates this field with with which I can manipulate the U scale and the the W scale in this example. V is to the top. I’m not caring about this because we just have one one pin. But you could also use that actually to have bowing the pins for example, but just to keep it simple as a as a starting point and proof of concept, right?

17:26 I I chose this and we’re also just looking at n equals two, so eight optimization parameters that will will be four for the U scale and four for the V scale. All right, so optimization as I said, I wanted to keep it quite simple to start with to to really yeah, prove the concept. So, we’re only have eight optimization parameters. I only ran it for for 20 iterations.

18:00 I also didn’t really put any any constraints. The one thing I wanted to do is I wanted to make sure that the pins actually don’t completely overlap, which would also be possible if the parameters get go to very high values. So, I I limited the the parameter space that I’m actually employing to make sure that the pins don’t overlap. And just one simple objective, I’m minimizing the the solid mean temperature.

18:28 The beauty of this is I’m really just running it on my laptop. So, I’m running it on a on a laptop GPU with 12 GB of of a VRAM. And the average simulation time as you can see in the picture on the bottom left is 60 seconds. Well, more precisely you can basically see there are some simulations take 40, around 40, others 60 and then there’s three that are outliers that take much longer.

18:54 This is basically when the fluid solver tells me that it diverged and it reduces the time step, calculates again and this is how we end up with a higher simulation time, obviously. In the right hand side, I’m plotting the temperatures that that, the the the mean temperatures that I get back from the configurations that were investigated by the by the optimizer. You can see the baseline starts at around 925 Kelvin and the best iteration in this limited set up to 20 iterations was iteration 15 with a minimum temperature of 867.9 Kelvin.

19:40 So, that’s quite a significant reduction. And all this is a result that I could obtain within an hour. So, this is something I could just start before going on lunch break, come back and already have an idea of how I could have a more optimal pin fin bank in my design. I think that’s that’s pretty impressive. That’s definitely something I’ve been always dreaming dreaming of. So, let’s take a a quick look also at the at the results to understand a little bit where does the reduction in temperature actually come from.

20:21 I mean, it’s it’s pretty wild, but I I hope you can you can see that compared to here in the baseline we have a lot more red areas. And this all goes rather to the yellow and and orange range. Also, when we cut it open, you can also see on the pins directly that all the temperatures even all the way to the trailing edge basically go down and there’s some things that that help here, right?

20:47 Like this is exactly one of the phenomena I was talking about. Like if you assume that the the flow either goes like this or like this through the pin fin array, this is not what’s going to happen, right? Like if it sees the possibility to go 45° like direct line of sight to an exit, it will do that, right? And that’s very often not what you want.

21:11 And you can see by putting more pins here, by increasing the pin density in this area, this is exactly something that the optimizer then avoids. And in contrast, it it now benefits the flow to to to pass through these pins here, where it’s a little bit more closer. That increases the speed, increases the heat transfer coefficient, and therefore the temperature goes down. I mean, full disclosure, if you look at the at the velocities here in the streamline picture at the outlet, you can see that the optimized result actually creates much higher discrepancy between the slots, the velocity you have in the slots.

21:56 That’s definitely also not ideal, right? You can basically see that here, that now we have a lot of very cold slots next to quite hot slots, and that is also a problem that just tells me I should have set up my optimization example also a little bit differently, because obviously it’s not only just the average solid temperature that plays a role, but I also need to look at temperature gradient.

22:22 I also need to look at the maximum temperature that cannot go be above a certain value, and I also need to make sure that I don’t have in critical geometric locations or collect the trailing edge at too strong temperature gradient. So, to sum it all up, basically we ran this successful proof of concept. There’s some additional validation for sure needed to really also look into the flow results, whether they make sense compared with other simulations, but the geometric robustness and simulation speed was really quite quite impressive.

23:01 And And this is exactly what we will need to early on during the design already I use the kind of of the simulation to really guide our design. Obviously, there’s a lot of work to be done with appropriate objective and constraints. And then for potential objectives, what I said in the beginning, here I’m just basically looking at the at the thermal performance, but what about pressure loss?

23:28 What about structural integrity? What about manufacturability and all these aspects that need also to be considered? Could increase the parameters, just the iteration count. There’s lots of things that we can do. But what I have already started is to go from a really a sub model now rather to look at the whole the whole blade because one of the first things we as designers need to decide is where do we actually place the ribs.

And that is not a just a geometric task, but that really depends on how we want to how we want to guide the flow and have a have a good flow distribution. So, the model is already set up and that’s probably the the problem I’m going to try next. I hope that was interesting to you. If you have any question to the work that that we’ve done, how Ansys Fluid works, more interest in the example that I was presenting, here’s our contact information.

24:30 Talk to us now or just shoot us a mail. And then, thank you very much for your attention. To learn more about the CD-adapco Computational Design Symposium, access the archive of previous presentations, interviews with speakers, and information about future events around the world, visit CDFAM.com.