Altair Antenna Design Optimization with Machine Learning
Altair, Troy, Mich.

Altair’s simulation-driven approach to innovation is powered by an integrated suite of software which optimizes design performance across multiple disciplines. Encompassing structures, motion, fluids, thermal management, electromagnetics, system modeling and embedded systems, the suite also provides data analytics and true-to-life visualization and rendering. Altair’s vision is to transform decision making by applying simulation, data analytics and high performance computing.

Altair has been pioneering artificial intelligence (AI) and machine learning (ML) using their design exploration (DoE) tool, Altair HyperStudy™, for many years. HyperStudy automatically creates intelligent design variants, manages runs and collects data. Users are then guided to understand data trends, perform trade-off studies and optimize design performance and reliability. Taking advantage of design exploration methodology in HyperStudy, ML has been incorporated for antenna design optimization and electromagnetic compatibility problems using electromagnetic simulation tool, Altair Feko. ML approaches are applied to complex antenna designs with many design variables. The complete workflow of the ML approach for antenna design optimization is detailed in these steps:

  • Generate training and test data with an appropriate DoE study and numerical simulation
  • Build a ML model based on the generated training data
  • Validate the ML model using the generated test data
  • Use ML model to optimize antenna design

Antenna design with ML helps in understanding data trends, perform trade-off studies and optimize design performance and reliability. Two examples are provided to illustrate advantages of applying ML for antenna design process using Feko and HyperStudy.

Figure 1

Figure 1 Optimization of dual port LTE antenna.

Figure 2

Figure 2 Optimization of WLAN antenna with evolutionary learning.

Example 1: Dual port LTE antenna (see Figure 1)

  • Antenna simulation model with 12 design variables and two responses (S11, S21)
  • Use ML regression model for fast optimization. Comparison of processing times:

Data generation with DoE: 14.7 hrs

Training of ML regression model: 2 sec

Optimization based on ML model: 10 sec

Regular optimization based on simulation model: 40.4 hrs

Speed up factor 2.75 by using ML approach

Example 2: Optimization of WLAN antenna with evolutionary learning (see Figure 2)

  • Antenna simulation model with 112 (binary) design variables and two responses
  • Design variable indicates if honeycomb element is metallic or not
  • Number of possible design combinations: 2112 ≈ 5.2 ∙ 1,033
  • Find optimal antenna topology using evolutionary learning
  • Optimization goal: Minimize S11 at 2.44 and 5.2 GHz
  • Optimization constraint: Sum of metallic honeycomb elements
    < 50
  • After 4,300 iterations and 12 generations the multi-objective genetic algorithm has identified a set of Pareto-optimal solutions

Altair has been using regression type ML algorithms for design exploration and optimization for many years and now also applies this approach to antenna and related electromagnetic problems. Altair has recently expanded its ML portfolio with new products that address data analytics to include ML and predictive analytics. Application of Altair’s modern ML tools to antenna design will revolutionize development of new and innovative antennas for current and future wireless devices and products.

Machine Learning with Ansys Physics-Based Simulation
Ansys Inc., Canonsburg, Pa.

Ansys has embraced and used ML methods and tools for quite some time with the goal of advancing the capabilities of physics-based engineering simulation. An early example is the automated methods that Ansys HFSS uses during the adaptive meshing solution process. The method uses previous finite element solutions to predict where more elements should be placed for more accurate solutions. Although not ML in the modern sense, making predictions or decisions without being explicitly programmed to do so, automated adaptive meshing leverages the computer to perform the work of finding an optimal mesh for accuracy and speed based on previous observations.

In semiconductors, the RedHawk-SC product leverages big data and ML to enable rapid design iterations of exceedingly large and complex integrated circuit designs. ML provides actionable analytics to identify and prioritize design fixes. Ansys is also making use of ML methods in specific simulation capabilities such as inference of optical properties of materials, designing smart assistants for gauging high performance computing (HPC) resource usage prior to simulation and automatic road scenario generation for advanced driver-assistance systems and automotive radar.



Figure 3 Typical range-Doppler plot.

Ansys has demonstrated the use of the Ansys HFSS Shooting and Bouncing Ray (SBR+) asymptotic electromagnetic solver to predict radar returns in a complex driving scenario. Ansys HFSS SBR+ is used to predict the Doppler response for a vehicle-mounted radar as it travels through a moving scene that includes the road and combinations of stationary, moving vehicles and persons, buildings, road signs and foliage.

Using that capability, the software may apply ML algorithms to provide radar-based object localization and classification. In ML, the goal is to detect and infer certain patterns from complex data by generating a mathematical model that can be used to quickly make decisions on new data. It is generally rare for pre-trained models or datasets to exist that can be utilized in ML training; thus, the method of this work is to generate the datasets and training using the Ansys HFSS SBR+ computation.

Advantages of using this simulation-based approach is that the data may be generated automatically using any driving scenario that may not be possible empirically. More importantly, the objects in the scene were placed there when creating the scenario providing the unique ability to self-annotate and label the data. We know if there is another automobile or pedestrian because we placed it there. This is in significant difference to empirical methods that require massive and painstaking hand annotation. Of course, dataset creation is accelerated using HPC especially considering the highly parallel simulation requirement.

Velocity of an object is effectively determined by Doppler shift in the returned signal and distance (range) from the radar is effectively determined by time delay. A typical radar return (see Figure 3) provides magnitude of scattered fields in a color shaded image, with range and range rate indicated in the plot via location along the Range and Doppler axis, respectively. Everything in the environment scatters fields so automobiles, pedestrians, signs, buildings and trees all contribute and may show up in radar image.