Automatic Cooling Circuit Generation for Injection Molding

DAVID ASTBURY: Welcome to the session today, where Clinton Kietzmann and myself are going to take you through some of the work we've done over the past 12 months on the automatic cooling circuit generation for injection molding. This project started as an industrial partnership aimed at developing a generative solution for conformal cooling for plastic injection molding. It's been a partnership that's running about two years now between Autodesk and Panasonic. And the real objective is to try and make the design of optimal cooling circuits a lot less mysterious and trying to evolve really high performance cooling circuits with not a lot of knowledge from either a CA analyst or engineer.

So the objectives are really to improve part cycle time and part quality. This is an extension of the work that was presented last year at AU. There was a paper presented on generative design for mold cooling channels. And you can see some of the outcomes there, some of the different cooling circuits that were generated for that part of the project. Some limitations that we did find with that project was that for the approach that was being used, which was a Voxel solver, it was great for analyzing just inserts. But it didn't really scale to the whole mold. So this year we've changed our approach. And by using traditional flow solvers, we're attempting to look at the optimization of a cooling channel for a whole mold.

OK before we start let's think about, well, why does cooling matter. Basically, the main reason you want to have a good cooling circuit is it will increase the profitability of your part. And it does that in two ways. The first way is anything that reduces your cycle time means you can produce the part much more cost effectively or quicker. And the other aspect is part quality. If you have a hole for cooling or unoptimized cooling, you can get hot spots and cold spots in your mold. And that can contribute to things like warpage, increase your defect rates, and also increase your cycle time. So those two aspects are quite important in determining how profitable a tool is. OK?

There is a relationship between cycle time and product quality. They are linked, but it's in a very complex way. And generally, you can have one or the other. Only in optimal layouts can you generally get both. So you see some sort of compromise you'll need to make between the cycle time and the part quality.

OK. So let's get on to looking at the blueprints in terms of what you need to do to generate cooling circuits automatically. OK, there's really three parts to this. What we're trying to do is determine the optimal locations and geometric properties for cooling channels to maximize our part quality and minimize our production and tooling costs whilst respecting constraints due to the original part design, cavity layouts, mold construction method, mold tooling elements, and ancillary equipment. And I'll spend the next few slides just breaking down each of those three areas into what it really means.

OK. In terms of the scope of this project, we're going to reduce the design and optimization degrees of freedom. We're going to assume that the coolant inlet and outlet are in fixed locations and provided. We're going to assume that the coolant inlet temperature and coolant flow rates are constant. We're going to assume that the cooling circuits are circular, and have constant diameter, and there's no branching or bubbles and baffles. The algorithms that we have developed will account for those things. But in the time we have available, really what we're trying to do is just show the concept of how producing an optimal circuit layout gives a lot of benefit. And the last constraint here is we will be assuming that the model is 3D printed. And that's to allow us more degrees of freedom.

So if we look at what's in scope, then, really what we're trying to do is change location, the position, and the path of a circuit. So if we look at the case on the left there, that's probably the simplest circuit that we can have. Two circuits, one in the fixed half, one in the moving half. And you can see there's quite a bit of temperature variation across the cavity there. And by changing location, position, and path we can reduce that to a certain degree, and produce a more optimal design as seen on the right hand picture there.

So in terms of quality, what do we really mean? One of the quality measures we're trying to look at is the uniform temperature distribution across the part's surface. And if we minimize the part's surface temperature variation, we're going to get a higher quality part. Now the measure we use for this is the standard deviation of all the part's surface temperatures. So if we have every element, or every part, of the surface of the part being at the same temperature, then we'll have our optimal quality.

Now why is that important? If we have a temperature difference, then that can lead to volumetric shrinkage difference, which can lead to increased warpage. Also, if we have different temperatures on the surface of a part, it can lead to different gloss levels and visual defects on the surface of the part.

So if we look at a typical cavity, there's certain areas in the cavity where it's going to be difficult to keep the variation of the surface temperature small. If we look at things like ribs, if we look at the bottom of the rib we have effectively heat coming in into the corners of the ribs, so there's much more heat being put into a small area of the tool. And we'll end up with hotter areas in there. And this is why we need to get cooling into those areas, or maybe even use things like bubbles and baffles to help extract the heat in those areas.

If we look at edges, edges have the opposite effect, in that they're typically cooled from multiple sides so they tend to cool a lot quicker than the rest of the part, and that can introduce a temperature variation that way.

Bosses, generally quite thick and chunky, and therefore are going to take a lot of time to cool. Thickness change, everywhere we have a different thickness, it'll take a different amount of time to cool. So we really want to try and make sure that there's not a lot of surface temperature difference across our thickness changes.

Injection locations can cause problems because one, they're the greatest source of heat. There's a lot more heat gets introduced into the mold around where the injection point is. And also they create an obstacle which limits some of the places we can put our cooling. And the last one there is pockets or small thin areas where we really can't get any cooling into the cavity. And they can get quite hot, and can have a significant impact on the quality of the parts. So there's quite a number of areas within the cavity that need to be considered very carefully when locating cooling channels.

In terms of minimizing our cycle time, the strategy we're trying to employ here is to put our circuits as close as possible to the part to extract the heat as quickly as possible. And if we do that, that will improve our cycle time. So the measure that we're looking at here is area weighted average of the part surface temperature. So if we increase the-- sorry, if we decrease-- the average temperature, then we extract from the heat out of the part quicker, then run cycle time will decrease. OK.

Another aspect with respect to the tooling that we do need to consider, is that spending more time on an optimal cooling layout may be more cost effective over the life of tool. It might be a lot more expensive to create the tool. But if we can reduce the cycle time by two, three, or five seconds over the life of the tool, it may well be a good payoff. And another thing worth noting is there that cooling circuits that follow the shape of the part generally form better. The so-called conformal cooling circuits. They may be more expensive to build, but again, over the lifetime of the tool, they may help get our average temperature down and improve our cycle time, therefore being more inexpensive or cheaper to produce our parts, or we can produce our parts quicker.

So if we look at the impact of manufacturing method, if we consider a simple cavity here, using the drill and plug method on the left, we have just two simple circuits, our starting circuits, and part with a slight curve in it. To get the heat out as quickly as possible, we want to get the cooling circuits as close as possible to the part. So using a drill and plug method, that means that we're going to have to create bends or junctions or multiple drill holes to get those circuits as close as possible to the part. And that can be quite expensive.

If we look at doing the same sort of thing with 3D printing, effectively we can move any of the nodes or joints in the system. We can really get them quite close to the part and give us a much more optimal cooling pattern. And there's very little cost difference in printing with 3D, regardless of the cooling circuit layout that you choose.

So just to summarize how the manufacturing method can impact things, with drill and plug, the cost increases with every channel. With 3D printing, the cost is independent of the number of channels. With drill and plug, it may be quite difficult to introduce conformal cooling, whereas if you're using 3D printing it comes for free. Drill and plug may have better lifetime in the tools. Obviously 3D printed tools have a shorter tool life and issues of things like fouling of the cooling channels can be more common. With drill and plug, it's possible to add inserts. With 3D printing, some of the latest 3D printers do allow you to print with multiple mold materials. That opens up a whole heap of possibilities where you can graduate from one material to another and change the conductivity of the mold to redirect the heat pattern. And that's an area, I think, we'll be seeing a lot more in the next few years.

With the drill and plug you can support bubbles and baffles quite easy. With the 3D printing, you can't directly print a bubble or a baffle, but you can print your 3D circuit and then, in a post-processing operation, introduce bubbles and baffles into the tool. I think that's sort of the typical way things will go, maybe, in the future where there'll be the combination of 3D printing and other operations, afterwards, to put in devices such as 3D bubbles and baffles.

OK. In terms of the tool, there are many constraints. Within the actual tool itself, cooling circuits need to keep clear of the following feed systems, the ejection system. Generally you would put your ejection system in first and design your cooling channels around those, because if you can't eject your part, then that's going to create all sorts of problems. Core pins, mold edges. Obviously you need to keep a little bit of steel between the outer boundary of the mold and the cooling channel and similar part services, and to other cooling circuits and vents.

So really what we're suggesting is that a cooling circuit needs to be at least one times its diameter from any of the other mold components. So if a circuit was 5 millimeters, then you have to have at least 5 millimeters of steel around that. If it was 10 millimeters, at least 10 millimeters of steel around that. And that's to ensure that you don't have any issues, and to ensure that the structural integrity of the mold and mold assemblies is maintained.

The other constraint that we typically see is when you have multi cavities. Unless the system is completely symmetrical in terms of the cavities and the cooling, then you can get some cavity to cavity interactions which may lead to part to part variations. So if we look at the system here at the bottom right picture there, We can see that there's a 12 cavity system there. And we can see that the coldest cavity is the bottom left one, and the hottest cavity is the one that's on the top, second from the left. And the reason that cavity is hotter is because there's an interaction between one cavity and the other.

If we look at the top. left cavity, we can see it gets cool to the mold block on two sides, whereas the second from the left only gets cooled to the mold cavity from the top. And also the cooling channel that it's working with has a temperature rise across it. So in this case, there's a 5.2 degree temperature rise across the circuit. So that can lead to different volumetric shrinkage, and different part to part variation. OK? So in general, when you're designing a circuit, you really want to keep the coolant rise as low as possible. And definitely not more than 10 degrees Celsius. So the advice here is, where possible, use a symmetric cavity layout, and a symmetrical cooling layout. Easy to say, quite difficult to do, given all the considerations of the things that we need to work around or the obstacles within the tool. But

The other constraint we need to look at is the coolant inlet. So for a cooling channel to be effective, we need to make sure that the inlet flow rate, or the flow rate in the circuits. Is greater than 10,000 to maintain turbulent flow. This means we need to have sufficient pump capacity to generate turbulent flow. So, the pressure drop in a circuit is proportional to the length of the circuit and the number of bends in a circuit. So as we make circuits longer and have more and more bends and curves, the pressure, or the pump capacity, to require that Reynolds number, will increase.

The other constraint in terms of the inlets is typically you find that the coolant inlets and outlets are on one side. One side of the machine, is normally a wet side of the machine, and a dry side of the machine. And in some cases, they may be fixed to allow things like quick change mold systems to work correctly.

OK. So, in general, what are we trying to do? , Effectively we're trying to create something from nothing. So if we define an inlet and an outlet, and we take into account the constraints imposed by the part design, the cavity layer, the inlet boundary conditions, really what we're trying to do is determine how do we generate an optimal cooling circuit that gives us the best performance? So if we look at the picture there, we can see there's an area in the fixed half and an area in the moving half where we can place cooling circuits. So how do we really determine where the circuits should be placed?

OK. So I'm going to outline an approach that's based on geometry. So the principle here is to optimize heat removal by placing cooling circuits as close to the part as possible using different geometric methods and predefined patterns. So the aim is by getting the cooling circuits as close as possible, we're going to give the shortest possible cycle time. So, how does this method work? What we're doing is effectively employing draping.

So if you look at the part on the left, there, what we do is identify a parting plane, or in the case of a multi-plate tool or a stack tool, we identify the parting planes that we want to work with. And what we then do is we treat each of those areas separately and apply a different circuit pattern. So if we look in the middle, in the top, we're applying a certain type of pattern. And in the bottom, we're applying a different type of pattern. And then effectively, what we do is we drape those patterns onto the part to get our cooling channels close to the part. This has a good advantage in the respect that it ensures that the cooling circuits reach all the external boundaries of our cavity.

So what are the strengths of this method? The primary strength is it ensures that we get cooling to the part extremities. Computation is inexpensive, so it's quite quick to do. There are known, good, patterns that exist based on experience. And it's easy to evaluate many different designs in a generative way. So we can generate a number of different patterns of quite fundamentally different shapes, and see what impact that they have on the cooling circuit. And we can also paramatize patterns to introduce different layouts as well.

So in terms of patterns, what type of patterns are we looking to employ? The simplest is a single circuit, where we can crisscross backwards and forwards across the part, and drape that over cavity. We can, as in the second case there, have multiple circuits and drape that onto the cavity. And we can even do much more complicated circuits, like the hexagonal mesh, or even a spiral pattern to investigate the impact of each of those.

We can also look at things like orientation, whether we orientate the cooling circuit in different directions relative to the part can make a big difference to the effectiveness of the cooling. And we can also change things like the spacing, how close the cooling circuits are spaced together, and other parameters around that area.

So if we look at some examples, if we look at the example on the left there, at the mouse, we can see that the circuits follow the contours of the parts quite well. We can also extend the circuits out towards the edge of the cavity to help with the heat dissipation as well. Because we're 3D printing this part, or the aim is to 3D print it, we can use the whole mold tool, and it really doesn't cost any more to produce a circuit that goes to the edge of the part, or one that may be, traditionally, you would have just had the cavities crisscross nearer to the edge of the part. In this case, it will also have a very long cooling circuit help with the heat removal from the circuit as well.

If we look at the bottom right, there, we can see another example of a fan blade. And again, we can see it's quite a complex structure. And we can see that the circuits are very closely conforming to the pattern of the shape of the part. We can also add variance, like we've had hoses in the end of this case. And another example in the top center there. When we look at this case, we can see that our injection point there on the left is that the circuits have actually gone and avoided the ejection point. So we've made an allowance for that. So that's a typical example of how we can use this sort of draping approach to make sure that all our part is being covered adequately by a cooling channel.

There are some limitations with this approach. If we look at the glass in the center, there we can see that there are some areas of the glass where there's quite a lot of cooling. And we can see areas on the side where there's slightly less cooling due to the nature of the starting shape. So in that case, this sort of pattern may be non-optimal.

The other thing we need to be aware of when we're designing these sorts of circuits is that if we produce very long circuits, it does increase the pressure drop and it does increase the chance of having quite a high coolant temperature rise across the circuit. So we might see variation from one side of the part to the other because it's hotter on one side of the glass than the other.

So how can we avoid things like this? Maybe for this type of part where we have a deep, deep vessel, we might be better to employ the use of spiral circuits rather than a draping. So you can see an example on the right there, where we've deployed a spiral circuit on the inside and the outside. And that may well give us a better temperature variation, or lower average temperature, because we're getting the circuits a lot closer to the part and a lot more even in their distribution.

The other main limitation of this approach, is that there's no consideration for the part temperature distribution. Effectively, we're draping every part of the cooling circuit to the same distance from the part. And some areas of the part might have more heat, and some areas have less. So that will naturally increase the amount of temperature variation we may get across the part. It will most likely give us the optimum cycle time, but it may introduce some quality aspects.

So, now I'll hand over to Clinton, who's going to talk a little bit more about how we can optimize these type of patterns, with an optimization algorithm to further improve the quality.

CLINTON KIETZMANN: OK. Good day everyone. Thank you, David, for taking the presentation to this point. And from now on, I'll go on about, let's talk about the optimization.

So with this method, what we do with optimization, we begin with an initial layout, which David's just spoken about. And from that user supplied, or generated geometrically solution, with a defined coolant inlets and outlets, we'll take that model into the existing cooling boundary element solver. And what I will describe is the result, or the routines, that actually move the circuits to their optimum positions, and how we go ahead with that.

All right. So we'll begin with our algorithm. Like we said before, it's based on a boundary element method. And the main advantage of using the boundary element method, is that as successive iterations of models occur, you don't have to remesh the mold. If we use the finite element method or another method, the mold volume-- as the cooling circuits change, the mild volume would have to be remeshed again. With a boundary element method, this is not required.

So then after every design iteration, or every new geometry we get after an optimization, a full cool boundary element analysis is performed, like every other mode flow user would use. There's no shortcuts or assumptions. It's the full solution that will be used. And then from the results from that previous iteration, the temperature results, the optimization routine concentrates on moving the cooling channels to the hotter areas of the mold. And that is the mold that is in contact with the part. So the surface areas of the mode that are in contact with the part that are hotter than the average temperature of the mold in contact with the part, the cooling circuits will kind of gravitate towards those points slowly.

So when that happens, the solver analyzes these results of each iteration. And then, as David has spoken previously, the compromises with defined metrics, so that we can put the compromises into numbers. And we will use these metrics to compare the different analyses with each other.

Just a note, that with the boundary element method, for each iteration, the boundary integrals need to be recalculated for the new channel positions. So the boundary element method is dependent on the geometry, where every part is relative to the other. And as we're moving the cooling channels, these relationships need to be calculated again. And then once they've been moved and recalculated, the solver will solve each and every channel position in a new study for.

So the best way to describe this optimization algorithm is a directed search method, which is trying to solve the conflicting requirements. So it's trying to minimize the average mold temperature in contact with a part, which David has spoken about before. And in order to do that, you want to move the cooling channels closer to the part. But at the same time, you're trying to minimize the temperature difference across the part's surface. And that is to reduce the warpage, because the closer you move the cooling channels to the part, you might get cold spots which will increase the temperature difference across the surface of the part. So for that requirement, to minimize the temperature difference, you generally want to move the circuits further away from the part. So there's a conflict there.

What we define the average mold temperature matrix, the t metric by the average temperature of the mold in contact with a part, for whichever iteration you're on, divided by the average mold temperature in contact with the part with that initial geometry that we started off with. And then obviously, if that metric is less than 1, then we know we indicating an improvement. There's an improvement in the solution. If it gets bigger than 1, we actually getting hotter, which is probably not a good sign.

Similarly, with the temperature difference metric, we take the standard deviation, which David spoke about, on every iteration, and divide it by the standard deviation of the initial geometry that we started off with. Then once again, if it is decreasing-- if it's less than one, and it is decreasing-- then it is an improvement on its previous iteration. And once it starts increasing-- once that metric starts increasing, then we've probably passed the optimal point in order to minimize the temperature difference across the part.

We will now look at how we use these metrics, so the evaluation of the model. So as we said before, it's a balance of conflicting requirements and then a weighted sum method is chosen in order to deal with the dual objectives. So the equation is given by the metric is equal to the alpha times the temperature metric plus 1 minus alpha times the standard deviation metric, where the weighting parameter, alpha, is chosen by the user.

So if we just look at the formula again, if the user chooses a parameter of alpha equal to 1, that will mean that the final metric will just be totally dependent on the average temperature in contact with the part metric on the left hand side. And the standard deviation metric-- which deals with the average temperature difference across the surface of the part-- that side of the equation will change to zero, and it will only be dependent on the surface temperature metric. And vise versa, if we choose the alpha to be 0, then the final result will be entirely dependent on the standard temperature difference across the surface of the part metric. So whatever alpha the user chooses, which we have as a default is 0.5. So it gives each metric equal weighting. The aim is to minimize that metric.

So how the solver works. The solver will run a set number of iterations. And it can stop once the metric is minimized, as per the user's chosen value of alpha. So the metric will decrease, decrease, decrease as the designs get better. And then once the cooling channels start moving too close to the part, and generally the standard deviation metric starts decreasing, the metric starts increasing. And we can stop the solution there, or it will just run for a number of iterations, and then the user will have to make his own judgments on that.

So now we'll just look at a few examples. So as we said before, the circuits move to the hotter areas of the part in the mold. So if you look at the model animation in the bottom left hand corner, you will see over there that the part starts off with four initial circuits that are quite far from the part. And the inside of the box is quite hot. So with every iteration, it moves closer and closer into the part. And you can see in that diagram how the temperatures inside the box region are cooling down as the circuits move closer.

With this routine, as David mentioned before, we need to maintain a specified minimum distance away from the part and other circuits. If you look at the picture in the center of the bottom, you'll see that the part is modeled as a transparent body. And you can see, as the cooling circuits move, you can see them move inside the box region. And from that graphic, you can see the top circuit and the circuits inside the box as they move closer, they maintain a one diameter distance from the part. That one diameter distance can also be set by the user. So that is a parameter that the user can set. But we use one diameter as the default.

We also need to maintain symmetry of the cooling channels. Having said that, cooling is never purely symmetric in cooling, because there's always a rise in the coolant temperature inside the channel. So you can never guarantee perfect symmetry. But in the way the channels move, we move them in a symmetrical way. The algorithms have been optimized so that they move in a symmetric way.

And then if you look at the bottom right hand corner, as you see how the cooling channels move, obviously the center of this model is exactly, the origin is in the middle of the part. So it is symmetric around either axis, x, y-axes. So you can see that the two cooling channels, the top two move inwards, and the bottom two move inwards, in more or less a symmetric way. And then if you look at it once, they are one diameter away from the part, inside outside. You can see that the cooling channels are also one diameter separated from each other in the final shape. This current example, using a metric of 0.5, which gives equal weighting to the temperature and the standard deviation, that showed a 45% improvement over the starting shape, which one can expect because the initial circuits were quite far away.

And then, we come to the collision avoidance. For this example, is ejector pins, but the collision avoidance also takes care of parting planes, the split lines of parting planes. So the cooling channels will not intersect a parting plane, and will maintain the user specified, in this case one diameter, distance away from the parting planes and the parts and everything else.

For this model, we have this microwave dish on the right, the microwave dish lid, which we use injection molding. And this model has ejector pins in. If you look at the bottom left diagram, or the bottom right diagram, you'll see a graphic of the ejector pins in the mold, which are used to the part. And from that you'll see the initial cooling circuits that were used. They were a circumferential circuit, both inside and outside the part.

And they are on the outside. There is a cooling circuit on the outside of the part. And you can see, as the algorithm works, the cooling circuits move in closer to the part. If you look at the bottom left hand diagram. And if you look at the bottom center diagram, you can see how the circuits move inwards. And if you compare the bottom center diagram with the top center diagram, that is the equivalent plot if now ejector pins were used. So you can have a look at that one, the bottom one, and see how the ejector pins influence the final position of the cooling channels, as opposed to if there were no ejector pins on the one on the top.

If we look at the table in the top right hand corner, you will see that the first iteration has the metric one. So that is the initial condition, the initial circuits that we began with. And then you can see for 11 iterations, we've got the table in the red box. That is the optimum. If you have a look, the number reduced from 1 all the way down to 0.681. So that is the optimum position. The next iteration, the matrix starts increasing 0.682, 0.683. And so the model in the red box is the optimum one, where the analyses finished.

And if we just have a look at the table, you can see that the average mold temperature in contact with the part reduced from 49 degrees to 41 degrees, and the standard deviation reduced from 5.718 to 2.9987. If we look at the table for this model, we only ran 15 iterations, and stopped it. If you look at the temperature metric, which is the second column, you can see that the metric for the average mold temperature in contact with the part, is still reducing. After 15 iterations it still hasn't reached the minimum one. However the standard deviation metric actually bottomed out or optimized at 0.523, which was one iteration before the optimum. And from that point on, it started increasing. So that just tells you that if you move the cooling channels closer from that point, you're actually going to start getting severe hot spots and cold spots in the mold. So for the 0.5 mix, the model at 0.681 is the optimum.

And then finally, if we just look at the slide, in the bottom right hand corner you can see how the cooling channels move in the z-axis, how the bottom cooling channels move up slightly toward the part but the top cooling channels where the heat is more concentrated move Downward at a much, much faster pace, and how the ejector pins get avoided as it moves down. So they move down into the hollowed out section of the microwave dish lid.

The next slide will just look at the surface temperature of the mold in contact with a part on the moving half comparison. So this is the hotter part of the mold, traditionally this is the hotter part. If we can see there, it's got a red-hot section in the middle. And if we look at the approach from the one end to the other, you can see that temperatures go from 53 to almost 60 degrees at the same time. And the optimized version is the exact plot, but on the optimized model on the right. And you can see that they go from 47 degrees on the outer circle. And the center of the circle's actually 44 degrees, which is actually cooler than the center of the original model, the unoptimized one on that side. So, if you have a look, you can see that the optimized one is much, much cooler.

So if we look at the bottom, like we said before, it went from an initial temperature-- a mean of 49 degrees-- to 41 degrees on the right. The standard deviation went from 5.718 to 2.99. And also, the range, the temperature range, was 19 degrees, 19.285 on the original model. But the optimized one, the entire range even dropped from 15.072. The standard deviation accounts for outliers which would be captured in the range.

So just to summarize, the average temperature dropped by 7.956 degrees. And the optimization of the standard deviation dropped by 2.721.

This is the same picture but of the fixed half, the surface temperature of the fixed half comparison. And once again, you can see that the optimized one on the right is much, much cooler. The temperatures are much closer together than they were in the other one. So once again, the average temperature, same as the previous one dropped by 7.956, and the standard deviation by the same.

And I think it's at this point, I hand the presentation back to David.

DAVID ASTBURY: OK. So to summarize, from what you can see from the results, the automatic generation of cooling circuits is practical and possible. By employing geometric techniques, such as draping, we can get a good coverage of cooling channels over our part. And by using the optimization that Clinton described, we can refine that to produce an optimal result.

The second conclusion is really determining an optimal cooling circuit location is quite complex, with many design constraints and limitations which need to be considered. And if you think that we're just considering this paper the effect of moving the channel itself, there's extra complexity that can be introduced when we think about the changes you could make to the inlets. You could make changes to the geometry of the circuits, change the thickness of the circuit as it went through the cavity. You can look at things like introducing bubbles and baffles and all those things. So there's a lot more complexity that can be introduced. But in general, we find that if you get the cooling circuit layout correct, then you will have a better performing cooling circuit.

There's further research that will continue in this area. We have a lot more of the design options that we didn't talk about today to explore yet. And we'll continue to explore those with our industrial partner, Panasonic, over the next period. And maybe we'll present another paper next year on what we've found.

Thank you very much.