Accelerating Electromagnetic Simulations with Machine Learning: An Introduction to EMA3D Connect

A new consulting group at EMA is addressing some of the biggest challenges in electromagnetic (EM) simulation.

We have named the group EMA3D Connect. The team supports your projects by using your data to develop machine learning (ML) models resulting in streamlined workflows and enhanced decision making.

“Machine learning for physics is a very new realm, and it requires a lot of domain expertise to get it right,” EMA Software Developer II Parker Hranicky said. “Here at EMA, regarding the realm of electromagnetics, we can build these machine learning models using our numerical simulations and domain knowledge to enable them to generalize as much as possible.”

Challenges of Traditional Workflows

Traditional workflows face a series of challenges such the cost of design optimization, the underutilization of data, and budget overruns. ML allows EMA to nullify these challenges by running simulations faster, enabling rapid iteration and design exploration.

“Doing interference on a model can take less than a second to minutes to predict exactly what a numerical solver would take hours or weeks to do,” Hranicky said.

Taking it a step further, ML allows EMA to integrate real data into a model through the lifecycle of an entire product. Benefits include real-time assessment cutting cost and time leading to faster product development and market success.

“It’s a feedback loop so you can constantly continue to improve your model with low fidelity, high fidelity, and then eventually incorporate lab and real data into it,” Hranicky said.

What is ML?

ML is a type of artificial intelligence (AI). It allows computers to learn and improve from experience without explicit programming or being fed large amounts of data. It allows computer systems to continuously adjust and enhance themselves as they learn more. As an ML system gets more data, it can improve more.

Advantages of ML are pattern recognition, automation, and continuous improvement. Nearly every industry uses ML. Tasks include optimizing shipping and delivery routes, managing inventory, automating factories, and securing organizations.

“I think the biggest value in machine learning is you can amortize the cost of generating the data,” Hranicky said. “Once you have the data and you have the model, it can really reduce that feedback loop for engineers.”

Introducing EMA3D Connect

EMA3D Connect has built an ML framework to accommodate unique customer requirements. This allows for complete model customization and deployment, enhancing your simulation capabilities. EMA3D Connect not only accelerates the design process but we can also develop digital twins for existing designs. With a focus on visualization, EMA3D Connect makes it easier for non-experts to better understand the results.

The EMA3D Connect execution plan is as follows:

• Generation and organization of simulation data
• Data pipelines for training preparation
• Building and training ML models
• Model deployment and version control
• 3D visualization of simulation and ML results

Data Generation and Pipeline

To start we need to generate data to train the ML system. EMA3D Connect works with EMA’s simulation software Ansys EMC Plus and Ansys Charge Plus. The software solutions utilize application programming interfaces (APIs) built on Ansys Discovery to train. These APIs configure and automate simulations for specific use cases. This allows us to change different variables such as illuminating angles, size of apertures, the width, length, and height of pieces. You can also use different types of enclosures. The system queues the configuration files into a cluster, which runs until it generates data.

“It’s a big challenge being able to generate high quality data that allows our models to predict accurately,” Hranicky said. “So, it’s very important, especially in machine learning for physics, to find clever ways to build prior information and inductive bias into these models so they can generalize better with less data.”

EMA3D Connect does this by using data-driven surrogate models to develop simulation models. Surrogate models are based on data symmetries. Figure 1 shows the symmetries available through EMA3D Connect. The Connect team looks at the data and chooses which surrogate model is best for the results you are looking for.

Building and Training ML Models

The processing framework EMA3D Connect has created does not work just for training ML models but it also works for the production environment. Preprocessors take raw data and transform it into a readable format for the ML model. Figure 2 illustrates how EMA3D Connect processes the data for both training and production.

During the preprocess stage, the system reworks a configuration file into a point cloud and eventually a matrix of numbers. This adjacency matrix creates graphs and enables GPU accelerated learning to take place.

“I think this is a really important part because if you don’t have the preprocessing in place, it can make everything challenging and messy to deal with,” Hranicky said.

Model Packaging and Deployment

Once EMA3D Connect trains and deploys models, you can package them in different ways. Currently, EMA3D Connect supports a few deployment options.

The first method is docker containerization, which packages a model in a Docker container for consistent deployment across environments.

“I think being able to dockerize and containerize the model or application is easy because it packages all of the dependencies that you need to make the model work,” Hranicky said. “So, all you have to do is run the container image to be to do inference on these models.”

The second method is by using a model registry. This technique pushes models to a registry enabling easy access and deployment. As more data generates, multiple versions of a model will emerge, much like a feedback loop.

“Once we have additional data, we can fine tune it on the new data and then push the new version and you can simply just pull the model from the model registry and use it,” Hranicky said.

We also support a hybrid cloud model.

Visualizing Results

There are several ways that EMA3D Connect can visualize results. We have maintained the input and output of these ML models in the same format used in EMC Plus and Charge Plus. This approach ensures consistency and compatibility across our systems, making it easier to integrate and analyze data effectively. We also can develop tailored applications that display model predictions and simulation results in an interactive format.

ParaView

We use ParaView for detailed 3D visualization and analysis. Figure 3 shows our ParaView plugin. This is an STL file with the 3D results of a radiated emissions demonstration.

“I think the real power of this is the filters that you get with ParaView are very powerful, so it allows you to import the results from our solvers and utilize the decades of work that has gone into building the algorithms behind filters and processing of your simulation results,” Hranicky said.

Users can also upload HDF5 files into ParaView. Hranicky says the plugin lazy loads time steps so you can load massive HDF5 files for 3D results. You also can upload CAD files and overlay it on the HDF5 file.

Omniverse

We leverage NVIDIA Omniverse for high-fidelity, collaborative 3D visualization, enhancing real-time interaction with model results. Figure 4 is an example of electrostatic discharge (ESD) results. EMA has a ParaView connector with Omniverse which allows for realistic photo renderings.

“It enables us to show how electromagnetic waves might propagate or there might be some arc on a PCB,” Hranicky said. “It really gives a visualization aspect to it to allow not just the engineers that design the models to know what’s going on, but to give the managers an idea as well.”

Transforming Simulation

ML is poised to transform science engineering. Hranicky says it can speed up numerical simulation by multiple orders of magnitude making things that were once impossible due to computations restraints possible. This includes generative geometry and design, optimization studies, and being able to evaluate new electromagnetic environments.

“There’s a lot of promise in this field and it’s very, very new,” Hranicky said. “There’s lots of research that is going into this, and I think it can transform the way that we do engineering design, reduce the feedback loop for design processes, and allow us to build the next generation of products and machines.”

Reach out to us here to learn more about EMA3D Connect and discover the possibilities with machine learning.