The architectural backbone of modeling and simulation is rapidly evolving to meet customer demand for highly accurate product designs generated with greater speed. The integration of deep learning—a subset of machine learning and artificial intelligence—enables simulation software to autonomously learn to solve complex design challenges, resulting in better products.
Deep learning heralds a new era of democratized simulation that will disrupt nearly every industry. This will allow almost anyone—even those with limited engineering training—to run simulations by providing a simple set of parameters and asking the software to simulate an object.
In-depth compatibility becoming mainstream over the next year, simulation tools will enhance the way engineers innovate new products, enabling them to deliver products faster than ever while reducing costs.
Improving Productivity and Product Design
When users run a simulation solver, there are about 1,000 different parameter settings they must choose from, ranging from defining the initial state to boundary conditions to set up the finite element solver. Once users have made their selections, they run the simulation and the software operates optimally.
Here lies the problem: engineers must determine which of the countless different parameter settings will help run the software most efficiently. The simulation will consume an enormous amount of time if they are unable to find the needle in the haystack.
To enhance productivity, embedded deep learning software engines will be trained to monitor and track how engineers use simulation software to solve a specific problem, such as the design of an aircraft wing. As the deep learning program observes a sample of a thousand engineers selecting parameter settings, the deep learning program will operate similarly, automatically determining every parameter for the simulation solvers. This simplifies the user experience by significantly reducing their workload.
Product Design Streamlining
Some software solvers often take 10,000 hours to operate. This means that users would be stuck waiting for more than a year while the software numerically solves the second-order partial differential equations at every mesh point for every finite element. But what if the simulation software could become smarter to run faster?
By integrating simulation software with deep learning capabilities, the software is educated when exposed to slightly different scenarios. This allows the software to deduce and process the simulation later using an educated guess to reach a solution. There are two approaches to accelerating simulation: data-driven deep neural networks (DNN) and physics-informed neural networks (PINN).
For example, after a user runs a computational fluid dynamics (CFD) simulation to measure airflow over a sphere, the simulation could also measure airflow over ten other shapes, such as an oval, square, triangle, rectangle, etc. These data-driven DNN models require a lot of training data to build a model.
Eventually, users will need to simulate an odd-shaped structure—such as a mobile phone—that is not one of the ten different geometries the CFD software has been trained on. The software will guess that the shape of the mobile phone is a rectangle and provide an output similar to that of a rectangle. Physics-informed PINN models require little training but rely on physics models to build the neural networks.
Ultimately, leveraging the power of deep learning inference capabilities, the simulation software will operate much faster. Instead of requiring 10,000 hours to run a simulation on a supercomputer, the solver can operate in a matter of minutes.
Revolutionizing Product Design
Traditionally, product designs are created with a human in the loop. The user generates a computer-aided design (CAD) drawing, then the software performs the meshing and solving. This is a lengthy process.
Generative design takes product design to the next level. Instead of a human in the loop being responsible for creating a CAD drawing and executing meshing and simulations, deep learning software will quickly create millions of different designs and present a few top choices for user consideration. For example, when designing an aircraft, a user can quickly visualize and inspect the designs offered by the software and then select the optimal one for prototyping.
Optimizing Future Product Designs
To monitor the efficiency of everything from field-deployed assets to factory machines, users have traditionally leveraged a digital twin—a virtual prototype of the complete system of a deployed asset—based on a data-driven artificial intelligence/machine learning model. But a digital twin of an asset is not enough. Users need a deep learning operational analysis mapped over the digitized twin data to better understand product performance and the accuracy of the operational modeling. This data can be used to optimize future product designs, simulation tools, and customer support.
Creating improved product designs will also be achieved by optimizing the simulation tool itself through Business Intelligence and Deep Learning. For example, a customer using a simulation tool from a provider over time might request that the provider create an improved, customized next-generation version of the tool to solve a specific problem. By receiving a collection of simulations from the customer along with their input into a deep learning model, the provider creates the optimized tool that allows the customer to develop enhanced product designs.
Pushing the Limits of Ingenuity
Integrating deep learning into simulation software promises significant benefits for users. From mirroring human decision-making to achieving human-like streamlining to integrating digital twin data, this next-generation software will greatly enhance product performance. Simulation software embedded with deep learning will soon open the floodgates of simulation design to almost anyone, accelerating and simplifying a product’s path to market.
