Solving Engineering Challenges with the Intersection of Simulation and AI

As artificial intelligence (AI) technology continues to advance, AI is being applied in numerous ways to revolutionize the way people live and work.   

For engineers, AI presents valuable opportunities to enhance the simulation processes used for design and development. Through simulation, engineers are able to better facilitate AI algorithms in applications, which has become crucial in the design process.  

By combining AI with simulation technology, engineers can address data quality challenges, replace complex high-fidelity simulations and improve the algorithms embedded in a wide range of application systems.  

Challenge 1: Improving Training Data for Models 

Projects are more likely to fail without robust data to help train a model, which makes data preparation a crucial step in the AI workflow. Engineers can find they are spending excess time trying to figure out why their model isn’t working when “bad” data is often the culprit. 

In fact, there has been a strong movement toward data-centric AI in recent years as the AI community recognizes a project is only as good as its data. Getting the training data right is vital to success. Many examples prove that it is far more beneficial to focus time on improving the training data instead of tweaking the AI model’s architecture and parameters, yielding much larger improvements in accuracy.   

However, collecting and cataloging real-world data is a difficult and time-consuming process. Another challenge is that AI models are static – they use fixed parameter values. Engineers have to remain cognizant of all the new data they will inevitably be exposed to that may not be captured in the training set.   

This is where simulation comes in. The use of simulation to augment existing training data has multiple benefits and helps engineers overcome all of the above challenges. Such benefits include:

1. Dramatically reduced costs. Computational simulation is less expensive than physical experiments.
2. Control over the environment – engineers can have full control over the environment and also simulate scenarios that are too difficult or dangerous to create in the real world.
3. Access to internal states that may not be measured in an experimental setup, which is very useful when debugging why an AI model doesn’t perform well in certain situations. 

Challenge 2: Approximating Complex and Expensive Systems with AI Models 

Simulation-based modeling of a system enables rapid design iteration when designing algorithms that interact with physical systems, such as an algorithm that controls a hydraulic valve. The problem is that this simulated system will require a high level of accuracy to effectively recreate the physical system with which the algorithm interacts.  

Historically, engineers would achieve this accuracy with high-fidelity models built from first principles, which take a very long time, especially for complex systems. The result is far less time for design iteration and not enough time for exploring better design alternatives. 

However, engineers can now approximate that physical system with a reduced-order model (an AI model), or even forgo the creation of a physics-based model and train the AI model from experimental data. A reduced-order model is much less computationally expensive than the first-principles model and allows the time and financial resources to iterate and evaluate more design options.   

There have also been other advances in the AI space, such as neural ordinary differential equations (ODEs), that combine AI training techniques with models that have physics-based principles embedded within them. These models can be useful when there are certain aspects of the physical system that the engineer wishes to retain, while approximating the rest of the system with a more data-centric approach.  

Challenge 3: Improving AI-based Algorithms  

Today, engineers are relying more on simulation when designing algorithms for application systems. Many methods are used, including developing virtual sensors, i.e., observers that attempt to calculate a value that isn’t directly measured from the available sensors. Linear models and Kalman filters are also used.   

However, the ability of these methods to capture the nonlinear behavior present in real-world systems is limited. Consequently, engineers are turning to AI-based approaches that are able to model such complexities. They use data to train an AI model that can predict the unobserved state from the observed states, and then integrate that AI model with the system.  

The AI model is then included as part of the controls algorithm that ends up on the physical hardware, which itself is limited with memory and performance. This restricts the types of machine learning models that are appropriate for such applications, and the engineers may need to try multiple models and compare trade-offs in accuracy and on-device performance.  

At the forefront of research in this area is reinforcement learning. Rather than learning just the estimator, reinforcement learning learns the entire control strategy. This has shown to be a powerful technique in some challenging applications, such as robotics and autonomous systems, but building such a model requires not only an accurate model of the environment (which may not be readily available), but also the computational power to run a large number of simulations.  

Looking Ahead   

AI and simulation offer the ability to develop, test and validate models before hardware is introduced. The implications for both cost and time are enormous. Industry tools have already empowered engineers to optimize their workflows and cut their development time by incorporating techniques such as synthetic data generation, reduced-order modeling and embedded AI algorithms for controls, signal processing, embedded vision and wireless applications.   

Without a doubt, as the complexity of systems and applications grows, AI and simulation are set to become even more essential to the field.