An Ansys optiSLang Overview and Optimization Example with Ansys Icepak

Ansys optiSLang is one of the newer pieces of software in the Ansys toolkit that was recently acquired along with the company Dynardo. Functionally, optiSLang provides a flexible top-level platform for all kinds of optimization. It is solver agnostic, in that as long as you can run a solver through batch files and produce text readable result files, you can use said solver with optiSLang. There are also some very convenient integrations with many of the Ansys toolkit solvers in addition to other popular programs. This includes AEDT, Workbench, LS-DYNA, Python, MATLAB, and Excel, among many others.

While the ultimate objective is often to simply minimize or maximize a system output according to a set of inputs, the complexity of the problem can increase dramatically by introducing constraints and multiple optimization goals. And of course, the more complicated the relationships between variables are, the harder it gets to adequately describe them for optimization purposes.

Much of what optiSLang can do is a result of fitting the input data to a Metamodel of Optimal Prognosis (MOP) which is a categorical description for the specific metamodels that optiSLang uses. A user can choose one of the included models (Polynomial, Moving Least Squares, and Ordinary Kriging), define their own model, and/or allow optiSLang to compare the resulting Coefficients of Prognosis (COP) from each model to choose the most appropriate approach.

The COP is calculated in a similar manner as the more common COD or R2 values, except that it is calculated through a cross-validation process where the data is partitioned into subsets that are each used only for the MOP calculation or the COP calculation, not both. For this reason, it is preferred as a measure for how effective the model is at predicting unknown data points, which is particularly valuable in this kind of MOP application.

This whole process really shows where optiSLang’s functionality shines: workflow automation. Not only does optiSLang intelligently select the metamodel based on its applicability to the data, but it can also apply an adaptive method for the improvement of the MOP. It will suggest an automatic sampling method based on the number of problem variables involved, which can then be applied towards refining either the global COP or the minimum local COP. The automation of this process means that once the user has linked optiSLang to a solver with appropriate inputs/outputs and defined the necessary run methodology for optimization, all that is left is to click a button and wait.

As an example of this, we will run through a test case that utilizes the ability to interface with Ansys EDT Icepak.

Figure 1: The EDT Icepak project model.

For our setup, we have a simple board mounted with bodies representative of 3 x 2 watt RAM modules, and 2 x 10 watt CPUs with attached heatsinks. The entire board is contained within an air enclosure, where boundary conditions are defined as walls with two parametrically positioned circular inlets/outlets. The inlet is a fixed mass flow rate surface and the outlet is a zero-pressure boundary. In our design, we permit the y and z coordinates for the inlet and outlet to vary, and we will be searching for the configuration that minimizes the resulting CPU and RAM temperatures.

The optiSLang process generally follows a series of drag-and-drop wizards. We start with the Solver Wizard which guides us through the options for which kind of solver is being used: text-based, direct integrations, or interfaces. In this case, the Icepak project is part of the AEDT interface, so optiSLang will identify any of the parameters defined within EDT as well as the resulting report definitions.  The Parametric Solver System created through the solver wizard then provides the interfacing required to adjust inputs while reading outputs as designs are tested and an MOP is generated.

Figure 2: Resulting block from the Solver wizard with parameters read in from EDT.

Once the parametric solver is defined, we drag and drop in a sensitivity wizard, which starts the AMOP study.  We will start with a total of 100 samples; 40 will be initial designs, and 60 will be across 3 stages of COP refinement with all parameter sets sampled according to the Advanced Latin Hypercube Sampling method.

Figure 3: Resulting block from the Sensitivity wizard with Advanced Latin Hypercube Sampling.

The results of individual runs are tabulated and viewable as the study is conducted, and at the conclusion, a description of the AMOP is provided with response surfaces, residual plots, and variable sensitivities. For instance, we can see that by using these first 100 samples, a decent metamodel with a COP of 90% is generated for the CPU temperature near the inlet. We also note that optiSLang has determined that none of the responses are sensitive to the ‘y’ position of the outlet, so this variable is automatically freed from further analysis.

Figure 4: MOP surface for the temperature of Chip1, resulting from the first round of sampling.

 If we decide that this CoP, or that from any of our other outputs, is not good enough for our purposes, optiSLang makes it very easy to add on to our study. All that is required is dragging and dropping a new sensitivity wizard onto our previous study, which will automatically load the previous results in as starting values. This makes a copy of and visually connects an output from the previous solver block to a new sensitivity analysis on the diagram, which we can then be adjusted independently.

For simplicity and demonstration’s sake, we will add on two more global refinement iterations of 50 samples each. By doing this and then excluding 8 of our 200 total samples that appear as outliers, our “Chip1” CoP can be improved to 97%.

Figure 5: A refined MOP generated by including a new Sensitivity wizard.

Now that we have an MOP of suitable predictive power for our outputs of interest, we can perform some fast optimization. By initially building an MOP based on the overall system behavior, we are now afforded some flexibility in our optimization criteria. As in the previous steps, all that is needed at this point is to drag and drop an optimization wizard onto our “AMOP Addition” system, and optiSLang will guide us through the options with recommendations based on the number of criteria and initial conditions.

In this case, we will define three optimization criteria for thoroughness: a sum of both chip temperatures, a sum of all RAM temperatures, and an average temperature rise from ambient for all components with double weighting applied to the chips. Following the default optimization settings, we end up with an evolutionary algorithm that iterates through 9300 samples in about 14 minutes – far and away faster than directly optimizing the Icepak project. What’s more, if we decide to adjust the optimization criteria, we’ll only need to rerun this ~14 minute evolutionary algorithm.

What we are most interested in for this example are the resulting Pareto fronts which give us a clear view of the tradeoffs between each of our optimization criteria. Each of the designs on this front can easily be selected through the interface, and their corresponding input parameters can be accessed.

Figure 6: Pareto front of the “Chipsum” and “TotalAve” optimization criteria.

Scanning through some of these designs also provides a very convenient way to identify which of our parameters are limiting the design criteria. Two distinct regions can be identified here: the left region is limited by how close we are allowing the inlet fan to be to the board, and the right region is limited by how close to the +xz corner of our domain the outlet vent can be placed. In a situation where these parameters were not physically constrained by geometry, this would be a good opportunity to consider relaxing parameter constraints to further improve our optimization criteria. 

As it is, we can now choose a design based on this Pareto front to verify with the full solver. After choosing a point in the middle of the “Limited by outlet ‘z’” zone, we find that our actual “ChipSum” is 73.33 vs. the predicted 72.78 and the actual “TotalAve” is 17.82 vs. the predicted 17.42. For this demonstration, we consider this small error as satisfactory, and a snapshot of the corresponding Icepak solution is shown below.

Figure 7: The Icepak solution of the final design. The inlet vent is aligned with the outlet side’s heatsink, and the outlet vent is in the corner nearest the heatsink. Primary flow through the far heatsink is maximized, while a strong recirculating flow is produced around the front heatsink.

The accuracy of these results is of course dependent not only on how thoroughly we constructed the MOP, but also the accuracy of the 3D solution; creating mesh definitions that remain consistently accurate through parameterized geometry changes can be particularly tricky. Though, with all of this considered, optiSLang provides a great environment for not only managing optimization studies, but displaying the results in such a way that you can gain an improved understanding of the interaction between input/output variables and their optimization criteria.

Ansys Sherlock: A Comprehensive Electronics Reliability Tool

As systems become more complex, the introduction and adoption of detailed Multiphysics / Multidomain tools is becoming more commonplace. Oftentimes, these tools serve as preprocessors and specialized interfaces for linking together other base level tools or models in a meaningful way. This is what Ansys Sherlock does for Circuit Card Assemblies (CCAs), with a heavy emphasis on product reliability through detailed life cycle definitions.

In an ideal scenario, the user will have already compiled a detailed ODB++ archive containing all the relevant model information. For Sherlock, this includes .odb files for each PCB layer, the silkscreens, component lists, component locations separated by top/bottom surface, drilled locations, solder mask maps, mounting points, and test points. This would provide the most streamlined experience from a CCA design through reliability analysis, though any of these components can be imported individually.

These definitions, in combination with an extensive library of package geometries, allow Sherlock to generate a 3D model consisting of components that can be checked against accepted parts lists and material properties. The inclusion of solder mask and silkscreen layers also makes for convenient spot-checking of component location and orientation. If any of these things deviate from the expected or if basic design variation and optimization studies need to be conducted, new components can be added and existing components can be removed, exchanged, or edited entirely within Sherlock.

Figure 1: Sherlock’s 2D layer viewer and editor. Each layer can be toggled on/off, and components can be rearranged.

While a few of the available analyses depend on just the component definitions and geometries (Part Validation, DFMEA, and CAF Failure), the rest are in some way connected to the concept of life cycle definitions. The overall life cycle can be organized into life phases, e.g. an operating phase, packaging phase, transport phase, or idle phase, which can then contain any number of unique event definitions. Sherlock provides support for vibration events (random and harmonic), mechanical shock events, and thermal events. At each level, these phases and events can be prescribed a total duration, cycle count, or duty cycle relative to their parent definition. On the Life Cycle definition itself, the total lifespan and accepted failure probability within that lifespan are defined for the generation of final reliability metrics.  Figure 1 demonstrates an example layout for a CCA that may be part of a vehicle system containing both high cycle fatigue thermal and vibration events, and low cycle fatigue shock events.

Figure 2: Product life cycles are broken down into life phases that contain life events. Each event is customizable through its duration, frequency, and profile.

The remaining analysis types can be divided into two categories: FEA and part specification-based. The FEA based tests function by generating a 3D model with detail and mesh criteria determined within Sherlock, which is then passed over to an Ansys Mechanical session for analysis. Sherlock provides quite a lot of customization on the pre-processing level; the menu options include different methods and resolutions for the PCB, explicit modeling of traces, and inclusion or exclusion of part leads, mechanical parts, and potting regions, among others.

Figure 3: Left shows the 3D model options, the middle shows part leads modeled, and right shows a populated board.

Each of the FEA tests, Random Vibration, Harmonic Vibration, Mechanical Shock, and Natural Frequency, correspond to an analysis block within Ansys Workbench. Once these simulations are completed, the results file is read back into Sherlock, and strain values for each component are extracted and applied to either Basquin or Coffin—Manson fatigue models as appropriate for each included life cycle event.

Part specification tests include Component Failure Analysis for electrolytic and ceramic capacitors, Semiconductor Wearout for semiconductor devices, and CTE mismatch issues for Plated Through-Hole and solder fatigue. These analyses are much more component-specific in the sense that an electrolytic capacitor has some completely different failure modes than a semiconductor device and including them allows for a broad range of physics to be accounted for across the CCA.

The result from each type of analysis is ultimately a life prediction for each component in terms of a failure probability curve alongside a time to failure estimate. The curves for every component are then combined into a life prediction for the entire CCA under one failure analysis.

Figure 4: Analysis results for Solder Fatigue including an overview for quantity of parts in each score range along with a detailed breakdown of score for each board component.

Taking it one step further, the results from each analysis are then combined into an overall life prediction for the CCA that encompasses all the defined life events. From Figure 5, we can see that the life prediction for this CCA does not quite meet its 5-year requirement, and that the most troublesome analyses are Solder Fatigue and PTH Fatigue. Since Sherlock makes it easy to identify these as problem areas, we could then iterate on this design by reexamining the severity or frequency of applied thermal cycles or adjusting some of the board material choices to minimize CTE mismatch.

Figure 5: Combined life predictions for all failure analyses and life events.

Sherlock’s convenience for defining life cycle phases and events, alongside the wide variety of component definitions and failure analyses available, really cement Sherlock’s role as a comprehensive electronics reliability tool. As in most analyses, the quality of the results is still dependent on the quality of the input, but all the checks and cross-validations for components vs life events that come along with Sherlock’s preprocessing toolset really assist with this, too.

ANSYS Discovery Live: A Focus on Topology Optimization

For those who are not already familiar with it, Discovery Live is a rapid design tool that shares the Discovery SpaceClaim environment. It is capable of near real-time simulation of basic structural, modal, fluid, electronic, and thermal problems. This is done through leveraging the computational power of a dedicated GPU, though because of the required speed it will necessarily have somewhat less fidelity than the corresponding full Ansys analyses. Even so, the ability to immediately see the effects of modifying, adding, or rearranging geometry through SpaceClaim’s operations provides a tremendous value to designers.

One of the most interesting features within Discovery Live is the ability to perform Topology Optimization for reducing the quantity of material in a design while maintaining optimal stiffness for a designated loading condition. This can be particularly appealing given the rapid adoption of 3D printing and other additive manufacturing techniques where reducing the total material used saves both time and material cost. These also allow the production of complex organic shapes that were not always feasible with more traditional techniques like milling.

With these things in mind, we have recently received requests to demonstrate Discovery Live’s capabilities and provide some training in its use, especially for topology optimization. Given that Discovery Live is amazingly straightforward in its application, this also seems like an ideal topic to expand on in blog form alongside our general Discovery Live workshops!

For this example, we have chosen to work with a generic “engine mount” geometry that was saved in .stp format. The overall dimensions are about 10 cm wide x 5 cm tall x 5 cm deep, and we assume it is made out of stainless steel (though this is not terribly important for this demonstration).

Figure 1: Starting engine mount geometry with fixed supports and a defined load.

The three bolt holes around the perimeter are fixed in position, as if they were firmly clamped to a surface, while a total load of 9069 N (-9000 N in X, 1000 N in Y, and 500 N in Z) is applied to the cylindrical surfaces on the front. From here, we simply tell Discovery Live that we would like to add a topology optimization calculation onto our structural analysis. This opens up the ability to specify a couple more options: the way we define how much material to remove and the amount of material around boundary conditions to preserve. For removing material, we can choose to either reduce the total volume by a percent of the original or to remove material until we reach a specific model volume. For the area around boundary conditions, this is an “inflation” length measured as a normal distance from these surfaces, easily visualizable when highlighting the condition on the solution tree.

Figure 2: Inflation zone shown around each fixed support and load surface.

Since I have already planned out what kind of comparisons I want to make in this analysis, I chose to set the final model volume to 30 cm3. After hitting the simulate button, we get to watch the optimization happen alongside a rough structural analysis. By default, we are provided with a result chart showing the model’s volume, which pretty quickly converges on our target volume. As with any analysis, the duration of this process is fairly sensitive to the fidelity specified, but with default settings this took all of 7 minutes and 50 seconds to complete on my desktop with a Quadro K4000.

Figure 3: Mid-optimization on the top, post-optimization on the bottom.

Once optimization is complete, there are several more operations that become available. In order to gain access to the optimized structure, we need to convert it into a model body. Both options for this result in faceted bodies with the click of a button located in the solution tree; the difference is just that the second has also had a smoothing operation applied to it. One or the other may be preferable, depending on your application.

Figure 4: Converting results to faceted geometry

Text Box: Figure 5: Faceted body post-optimization.

Figure 5: Faceted body post-optimization

Figure 6: Smoothed faceted body post-optimization

Though some rough stress calculations were made throughout the optimization process, the next step is typically a validation. Discovery Live makes this as a simple procedure as right-clicking on the optimized result in the solution tree and selecting the “Create Validation Solution” button. This essentially copies over the newly generated geometry into a new structural analysis while preserving the previously applied supports and loads. This allows for finer control over the fidelity of our validation, but still a very fast confirmation of our results. Using maximum fidelity on our faceted body, we find that the resulting maximum stress is about 360 MPa as compared to our unoptimized structure’s stress of 267 MPa, though of course our new material volume is less than half the original.

Figure 7: Optimized structure validation. Example surfaces that are untouched by optimization are boxed.

It may be that our final stress value is higher than what we find acceptable. At this point, it is important to note one of the limitations in version 2019R3: Discovery Live can only remove material from the original geometry, it does not add. What this means is that any surfaces remaining unchanged throughout the process are important in maintaining structural integrity for the specified load. So, if we really want to optimize our structure, we should start with additional material in these regions to allow for more optimization flexibility.

In this case, we can go back to our original engine mount model in Discovery Live and use the integrated SpaceClaim tools to thicken our backplate and expand the fillets around the load surfaces.

Figure 8: Modified engine mount geometry with a thicker backplate and larger fillets.

We can then run back through the same analysis, specifying the same target volume, to improve the performance of our final component. Indeed, we find that after optimizing back down to a material volume of 30 cm3, our new maximum stress has been decreased to 256 MPa. Keep in mind that this is very doable within Discovery Live, as the entire modification and simulation process can be done in <10 minutes for this model.

Figure 9: Validated results from the modified geometry post-optimization.

Of course, once a promising solution has been attained in Discovery Live, we should then export the model to run a more thorough analysis of in Ansys Mechanical, but hopefully, this provides a useful example of how to leverage this amazing tool!

One final comment is that while this example was performed in the 2019R3 version, 2020R1 has expanded Discovery Live’s optimization capability somewhat. Instead of only being allowed to specify a target volume or percent reduction, you can choose to allow a specified increase in structure compliance while minimizing the volume. In addition to this, there are a couple more knobs to turn for better control over the manufacturability of the result, such as specifying the maximum thickness of any region and preventing any internal overhangs in a specified direction. It is now also possible to link topology optimization to a general-purpose modal analysis, either on its own or coupled to a structural analysis. These continued improvements are great news for users, and we hope that even more features continue to roll out.

Icepak in Ansys Electronic Desktop – Why should you know about it?

The role of Ansys Electronics Desktop Icepak (hereafter referred to as Icepak, not to be confused with Classic Icepak) is in an interesting place. On the back end, it is a tremendously capable CFD solver through the use of the Ansys Fluent code. On the front end, it is an all-in-one pre and post processor that is streamlined for electronics thermal management, including the explicit simulation and effects of fluid convection. In this regard, Icepak can be thought of as a system level Multiphysics simulation tool.

One of the advantages of Icepak is in its interface consistency with the rest of the Electronic Desktop (EDT) products. This not only results in a slick modern appearance but also provides a very familiar environment for the electrical engineers and designers who typically use the other EDT tools. While they may not already be intimately familiar with the physics and setup process for CFD/thermal simulations, being able to follow a very similar workflow certainly lowers the barrier to entry for accessing useful results. Even if complete adoption by these users is not practical, this same environment can serve as a happy medium for collaboration with thermal and fluids experts.

Figure 1: AEDT Icepak interface. The same ribbon menus, project manager, history tree, and display window as other EDT products.

So, beyond these generalities, what does Icepak actually offer for an optimized user experience over other tools, and what kinds of problems/applications are best suited for it?

The first thing that comes to mind for both of these questions is a PCB with attached components. In a real-world environment, anyone that has looked at the inside of a computer is likely familiar with motherboards covered with all kinds of little chips and capacitors and often dominated by a CPU mounted with a heatsink and fan. In most cases, this motherboard is enclosed within some kind of box (a computer case) with vents/filters/fans on at least some of the sides to facilitate controlled airflow. This is an ideal scenario for Icepak. The geometries of the board and its components are typically well represented by rectangular prisms and cylinders, and the thermal management of the system is strongly related to the physics of conjugate heat transfer. For the case geometry, it may be more convenient to import this from a more comprehensive modeler like SpaceClaim and then take advantage of the tools built into Icepak to quickly process the important features.

Figure 2: A computer case with motherboard imported from SpaceClaim. The front and back have vents/fans while the side has a rectangular patterned grille.

For a CAD model like the one above, we may want to include some additional items like heatsinks, fan models, or simple PCB components. Icepak’s geometry tools include some very convenient parameterized functions for quickly constructing and positioning fans and heatsinks, in addition to the basic ability to create and manipulate simple volumes. There are also routines for extracting openings on surface, such as the rectangular vent arrays on the front and back as well as the patterned grille on the side. So, not only can you import detailed CAD from external sources, you can mix, match, and simplify it with Icepak’s geometry, which streamlines the entire design and setup process. For an experienced user, the above model can be prepared for a basic simulation within just a matter of minutes. The resulting configuration with an added heatsink, some RAM, and boundary conditions, could look something like this:

Figure 3: The model from Figure 2 after Icepak processing. Boundary conditions for the fans, vents, and grille have been defined. Icepak primitives have also been added in the form of a heatsink and RAM modules.

Monitor points can then assigned to surfaces or bodies as desired; chances are that for a simulation like this, temperature within the CPU is the most important. Additional temperature points for each RAM module or flow measurements for the fans and openings can also be defined. These points can all be tracked as the simulation proceeds to ensure that convergence is actually attained.

Figure 4: Monitoring chosen solution variables to ensure convergence.

For this simple system containing a 20 W CPU and 8 RAM modules at 2 W each, quite a few of our components are toasty and potentially problematic from a thermal standpoint.

Figure 5: Post-processing with Icepak. Temperature contours are overlaid with flow velocities to better understand the behavior of the system.

With the power of a simulation environment in Icepak at our fingertips, we can now play around with our design parameters to improve the thermal management of this system! Want to see what happens when you block the outlet vents? Easy, select and delete them! Want to use a more powerful fan or try a new material for the motherboard or heatsink? Just edit their properties in the history tree. Want to spin around the board or try changing the number of fins on the heatsink? Also straightforward, although you will have to remesh the model. While these are the kinds of things that are certainly possible in other tools, they are exceptionally easy to do within an all-in-one interface like Icepak.

The physics involved in this example are pretty standard: solid body conduction with conjugate heat transfer to a turbulent K-Omega fluid model. Where Icepak really shines is its ability to integrate with the other tools in the EDT environment. While we assumed that the motherboard was nothing more than a solid chunk of FR-4, this board could have been designed and simulated in detail with another tool like HFSS. The board, along with all of the power losses calculated during the HFSS analysis, could have then been directly imported into the Icepak project. This would allow for each layer to be modeled with its own spatially varying thermal properties according to trace locations as well as a very accurate spatial mapping of heat generation.

This is not at all to say that Icepak is limited to these kinds of PCB and CCA examples. These just often tend to be convenient to think about and relatively easy to geometrically represent. Using Fluent as the solver provides a lot of flexibility, and there are many more classes of problems that could be benefit from Icepak. On the low frequency side, electric motors are a good example of a problem where electronic and thermal behavior are intertwined. As voltage is applied to the windings, currents are induced and heat is generated. For larger motors, these currents, and consequently the associated thermal losses, can be significant. Maxwell is used to model the electronic side for these types of problems, where the results can then be easily brought into an Icepak simulation. I have gone through just such an example rotor/stator/winding motor assembly model in Maxwell, where I then copied everything into an Iecpak project to simulate the resulting steady temperature profile in a box of naturally convecting air.

Figure 6: An example half-motor that was solved in Maxwell as a magnetostatic problem and then copied over to Icepak for thermal analysis.

If it is found that better thermal management is needed, then extra features could then be added on the Icepak side as desired, such as a dedicated heatsink or external fan. Only the components with loads mapped over from Maxwell need to remain unmodified.

On the high frequency side, you may care about the performance of an antenna. HFSS can be used for the electromagnetic side, while Icepak can once again be brought in to analyze the thermal behavior. For high powered antenna, some components could very easily get hot enough for the material properties to appreciably change and for thermal radiation to become a dominant mode of heat transport. A 2-way automatic Icepak coupling is an excellent way to model this. Thermal modifiers may be defined for material properties in HFSS, and radiation is a supported physics model in Icepak. HFSS and Icepak can then be set up to alternately solve and automatically feed each other new loads and boundary conditions until a converged result is attained.

What all of this really comes down to is the question: how easy is it for the user to set up a model that will produce the information they need? For these kinds of electronics questions, I believe the answer for Icepak is “extraordinarily easy”. While functional on its own merit, Icepak really shines when it comes to the ease of coupling thermal management analysis with the EM family of tools.