The term “AI” has been freighted with meanings in the popular imagination long before realizing practical uses. Now that AI is finally showing its value, we can examine what it means in manufacturing and engineering, specifically for uses in simulation.
For simulation, AI typically entails machine learning and neural networks. Let’s look at three use cases for leveraging AI in simulation.
Reduced Order Modeling (ROM)
Reduced order modeling isn’t new. Traditionally, it uses a physics-based mathematical matrix to reduce the complexity of simulating a 3D system at the expense of fidelity while remaining physics-based. It can be a worthwhile tradeoff because the time savings more than make up for a small loss of accuracy.
Where AI comes in, specifically the neural networks and machine learning aspects of AI, is that it enables running a limited number of simulation variations to produce a machine learning model or “brain” that can then be applied to run further simulations significantly faster than before, even faster than physics-based ROMs. This allows engineers to quickly optimize a design. The types of simulations that can use this data-driven approach include acoustic, thermal and heat transfer, structural, and fatigue life.
System Level Modeling
System level modeling is mostly applied to 1D models to run simulations very quickly. AI is applied this way is called “system level learning” through an approach called deep reinforcement learning. It’s typically used to produce a control that works by interacting with the 1D simulation model (environment) to run hundreds of thousands or millions of fast 1D simulation sequences.
System level modeling is leveraged by AI in a self-play type of simulation environment that generates an optimal control mechanism. Often seen in controls engineering, system level modeling applications can include things like furnace temperature controls or other systems that use any control mechanisms. It is also increasingly being used on edge devices for industrial IoT applications given that thee resulting model has a truly small compute footprint, which means it can run in real time or near real time.
Generative Adversarial Networks
Generative adversarial networks (GAN) are used in the conceptual stages of design to model things that don’t yet exist. Generative adversarial networks pair a generator algorithm with a discriminator algorithm. In GAN simulations, the generator algorithm produces many fake new designs. The discriminator algorithm is then used to identify those designs with non-realistic results. By catching most of the fake (unworkable in the real world) designs, it can help designers discover new optimal design options not looked at before.
That said, to be successful, GAN simulations must be fed a lot of historical data from related applications. One current application for GAN simulation is Formula 1 racing. Racecar builders collect tons of data. That data can be used to run hundreds of virtual tests such as for computational fluid dynamics (CFD) to optimize the geometry for new airflow patterns. During the conceptual design phase, the historical data can be fed into a GAN to create realistic CFD models which can help guide the engineers into novel designs.
Thinking of taking the first step? Contact a Maya HTT expert today.