VALIDATION. Planar Transformer. TRFLO vs ANNSS #10

"It depends" is the most common answer among experienced engineers who shy away from saying which computer modeling software is the best as the context and details are unclear. What are your needs, how skilled are you, and what is the budget?

Further, legal and lock-in barriers often limit exposure to validation studies or genuine feedback. Using tools that are already well-known is a safe strategy, too. Unsurprisingly, most articles related to magnetics modeling use the same one.

In this article, we want to shine some "light of objectivity" and highlight a few differences between an open-source-based tool and a well-known, supposedly "the best of the best of the best" simulation tool in the industry.

Why?

Because there are many uprooted stereotypes, such as one tool is for industry, another for academics, and the 3rd one for nerds. Let's fix that.

While this article is our subjective, sarcastic, slightly unserious, and biased opinion, for those willing to dive deeper into the subject and get a more independent viewpoint, we uploaded the document that we left without changes for you.

What document?

We kindly asked a competent expert to make a model with two tools for educational purposes while documenting steps. We provided the TRAFOLO model and feedback and reached a good agreement on the results at the end.

Ultimately, you would be willing to learn how to model planar magnetics, right?

So, the REPORT and the TRAFOLO case.

The Study

The model is based on the transformer design from the APEC 2024 conference proceeding "Integrated Planar Magnetics Optimization for ? ? ? LLC Converter with Wide Output Voltage Range," by Abdulsamed Lordoglu, PhD et al. The article presents an optimized, integrated planar transformer design that employs 3D FEM simulations to address skin and proximity effects, calculate core losses, and handle gaps making it a complex and ideal case for planar transformer simulation.

Essential details about the model are that the frequency is 1MHz, and turns are made of 0.104mm thick copper. Skin depth is 0.065mm or almost half of the thickness. Winding width, however, is 5mm (secondary), 1.33mm (primary) and 0.6. mm (another primary), meaning that skin/proximity effects in the direction of winding width are substantial.

So, let’s fire up the simulations!

1st ITERATION - make all the inputs r?i?g?h?t? the same

As we got feedback, core losses were 2 times smaller in TRAFOLO, while winding losses were about 20% higher.

It's not what we expected, but we know from experience

the greater the difference, the easier it is to find the problem.

The first simulations are usually run on a coarse mesh (quick tests to ensure that current directions, magnetic fluxes, and other fields are physical, the solution has reached convergence, etc.), but that did not explain the significant difference in core losses, because mesh refinement usually has little impact on it.

Is the magnetic flux the same?

We were assured that the flux and frequency were about the same. Well, then, let's check the core loss calculation we suggested. These are equations

Does the Ke parameter bring x2 smaller loss in core loss?

The most significant errors come from the most straightforward assumptions.

It is well known that the industry relies on over a century-old simple yet easy-to-grasp equations and PDF graphs. While teams packed with talented students and professors compete for accuracy in the Princeton MagNet Challenge, we continue fitting complex P(B,f,T) data onto the 3, 4, or sometimes 5 coefficient models, which leads to substantial errors at wide frequencies or magnetic flux ranges.

Still, we could not think adding one coefficient would be such an accuracy killer until we got this graph.

Unbelievable! It turns out that we rely on different core loss data. We take those from OpenMagnetics as a P(B, f, 100°C) and then fit them into the 4-parameter model. Considering that there are at least 3 sources where you can get material data for 3F4:

1. Windows XP-only software
2. Steinmetz coefficient table
3. Princeton MagNet

It is mission impossible to conclude where to draw the line.

The author of the document obtained the graph by fitting curves. As we are neither core manufacturers nor have the equipment and skills to measure core losses, we rely on what is available, so take those core loss data with a grain of salt. No finger-pointing here.

We did not investigate this matter further as we were confirmed that the same core loss coefficients lead to almost identical core losses.

Perfect!

2nd ITERATION - check the mesh

After such great news, all the attention has shifted to the winding loss. Although 20% in many cases is already considered good enough because Rac/Rdc for such high-frequency planar transformers is challenging to calculate (even with Excel), we were eager to know the reason behind the difference.

Most FEM-based electromagnetic solvers are based on the same discretization models. Users should expect very close results. The only difference that sometimes makes the difference is in additional models. Those models are applied on top of numerical discretization to help resolve special cases such as wire types (homogenization models) or boundary conditions (e.g., permeability gap models). There might be some differences in element order or type, too. Still, considering that most electromagnetic solvers are set by default similarly, the user should expect a very close result.

The simplest solution is often the hardest to see

Before we reveal the reason why winding losses were different, let us mention two common misconceptions that early users make. They rely on automatic meshing and error values that the software requests or outputs.

I set automatic meshing and a 0.001 error. Isn't my result up to 0.1% accurate?

The error is usually relative, not absolute. Further, it is often related to the residual of the linear system and not to the error of core loss that you are calculating. We should not rely on the error that the numerical software claims and have to do some of the homework.

In short, the issue was related to the mesh being too coarse. From the report, we see that two refinement iterations followed, and in the end, the results were closer to what we had anticipated.

4th ITERATION - good enough

When we provided the TRAFOLO model, we had already done some mesh refinements. From experience, we can skip a few mesh refinement iterations just by refining the mesh in places such as

1. The inner faces of the planar transformer winding
2. Winding faces that are the closest to the gap

Usually, it takes 2-3 iterations and is a common practice in simulations. It is a very simple procedure when you run multiple identical simulation setups on different refinements until the difference in results is adequate.

After some time, we got good news that the winding loss in ANNSS also went up, and there is a good enough agreement of around a 6% difference.

There are more results to see in the document, but you got the point. The tool is as good as you can use it. While there might be more models and features in one, you might make the model and get a reasonable result in another.

Exploring design options with FEM

Once the model is verified and optimized (mesh, settings, etc.), the report author uses the model to investigate some design options. In this test case, the direction of the current in one limb was reversed (both primary and secondary), which affected the distribution of the magnetizing flux.

The findings showed that while copper losses remain relatively unchanged between both scenarios, core losses more than double when the current flows in the same direction.

Testing component designs with highly non-uniform current distributions in windings and magnetic flux behavior in the core is where FEM shines.

The final remark

Mr. Core, before starting TRAFOLO, was working as a consultant for a company that used ANNSS for simulations. It is a good software, but not perfect. That led Mr. Core and Mr. Choke to join forces to make something we would prefer to use ourselves.

However, we quite often get feedback that we need more validations. For this reason, we would like to ask our great readers for help.

Researchers should be those who try new tools, do validation studies, and help to improve them. The "Better safe than sorry" option wastes talent and often produces research of little attention.

With all the above said, we encourage aspiring researchers to consider using TRAFOLO or the underlying open-source ElmerFEM for their following studies. Particularly, suppose you're in the early stage of your research career. In that case, you should expect valuable feedback from us and the community and increased visibility references to your work.

We do not prohibit researchers from using our tool for comparative or competitive studies and are open to discussing differences. There is little that can't be fixed or improved. It just takes time and effort.

Try it yourself. Here is the REPORT and the TRAFOLO case.

Do you agree?

If yes, then try it yourself. Here is the REPORT and the TRAFOLO case.

If not, then leave a comment.

Acknowledgment:

We want to extend our gratitude to Abdulsamed Lordoglu, PhD for providing valuable feedback and data for the model setup in this article. It was a fascinating case, and we wish you continued success in your future work.