The beautiful intersection of simulation and AI

Simulation has emerged as a critical technology to help companies shorten time-to-market and reduce design costs. Engineers and researchers use simulation for a variety of applications, including:

• Using a virtual model (also known as a digital twin) to simulate and test their complex systems early and often in the design process.
• Maintain a digital thread with traceability through requirements, system architecture, component design, code, and testing.
• Expand their systems to perform predictive maintenance (PdM) and fault analysis.

Many organizations are improving their simulation capabilities by incorporating artificial intelligence (AI) into their model-based design. Historically, these two areas have been separate, but create significant value for engineers and researchers when used effectively together. The strengths and weaknesses of these technologies are perfectly aligned to help companies solve three primary challenges.

Challenge 1: Better training data for more accurate AI models with simulation

Simulation models can piece together real-world data that is difficult or expensive to collect into good, clean, and cataloged data. While most AI models work with fixed parameter values, they are constantly exposed to new data that may not be captured in the training set. If left undetected, these models will generate inaccurate insights or will fail outright, leading engineers to spend hours trying to figure out why the model isn’t working.

Simulation can help engineers overcome these challenges. Rather than adjusting the architecture and parameters of the AI ​​model, it has been shown that time spent improving the training data can often yield more comprehensive improvements in accuracy.

Because the performance of a model is so dependent on the quality of the data it is trained with, engineers can improve results with an iterative process of simulating data, updating an AI model, observing the conditions cannot predict well and collect more simulated data for that condition.

Challenge 2: AI for new features in the product

Simulation has become an essential part of the design process for engineers using embedded systems for applications such as control systems and signal processing. In many cases, these engineers develop virtual sensors, devices that calculate a value that is not directly measured by the available sensors. But the ability of these methods to capture the nonlinear behavior present in many real-world systems is limited, so engineers are turning to AI-based approaches that have the flexibility to model the complexity. They use data (measured or simulated) to train an AI model that can predict the unobserved state from the observed states and then integrate that AI model with the system.

In this case, the AI ​​model is part of the control algorithm that ends up on the physical hardware and usually needs to be programmed in a lower level language such as C/C++. These requirements may limit the types of machine learning models suitable for such applications, so technical professionals may need to try multiple models and make trade-offs for accuracy and on-device performance.

At the forefront of research in this area, Reinforcement Learning takes this approach further. Rather than just learning the estimator, learning reinforcement includes the entire control strategy. This technique has proven effective in some challenging applications, such as robotics and autonomous systems, but building this type of model requires an accurate model of the environment – never a guarantee – as well as massive computing power to run a large number of simulations.

Challenge 3: Balance between ‘right’ and ‘now’

Companies have always struggled with time to market. Organizations that push a faulty or faulty solution to customers risk irreparable damage to their brand, especially startups. The opposite is true, as “also-rans” struggle to gain traction in an established market. Simulations were a major design innovation when they were first introduced, but their steady improvement and ability to create realistic scenarios can slow down perfectionist engineers. Too often, organizations try to build “perfect” simulation models that take a long time to build, risking that the market has moved on.

To strike the right balance between speed and quality, technical professionals must recognize that there will always be environmental nuances that cannot be simulated. AI models should never be blindly trusted, even if they serve as approximations for complex, high-fidelity systems.

The future of AI for simulation

AI and simulation technologies have built and maintained their momentum separately for nearly a decade. Now engineers are starting to see a lot of value at their intersection, given the symbiotic nature of their strengths and weaknesses.

As models continue to serve increasingly complex applications, AI and simulation will become even more essential tools in the engineer’s toolbox. With the ability to develop, test and validate models in an accurate and affordable manner, these methodologies will only continue to grow in use.

Seth DeLand is product marketing manager data analytics at MathWorks.