Surrogate Model Simplifies Complex Physics Tasks

There's a problem with physical modeling: Engineers who need the results of those simulations often don't have time to wait. Add real settings with the required multiple independent calculations (for example, individual thermal, mechanical and electromagnetic elements in the system) and “multiphysics” calculations that are both realistic and real-time can seem like an either/or choice.

Programmers and modelers gathering this week in Burlington, Massachusetts, will explore new ways to develop multiphysics on the go in COMSOL Simulation Environment. Over three days of talks, workshops and demonstrations. KOMSOL users will consider new approaches to solve the problem of time constraints in modeling.

“Surrogate models is an exciting new technology where you take your full multiphysics model and compress it into a compact format that can be quickly evaluated using machine learning,” says Björn Sjödin, senior vice president of product management at the parent company in Stockholm, also called KOMSOL.

This problem is more common than just COMSOL. According to review published earlier this year in the magazine Procedia Computer ScienceA number of industries are experiencing modeling bottlenecks where, according to the authors, “performing high-fidelity modeling can even take weeks per project.”

Surrogate models, Came as the authors note, involve reducing the equations to simplified versions of larger modeling frameworks. In other words, surrogates capture the essential behavior of the specific systems being modeled, but without much emphasis. computational costs. Often this pruning process may involve strategically sampling the original complex model at key points, and then learning a faster approximation that can predict outcomes for new scenarios.

“You can evaluate these models instantly,” Sjödin says of COMSOL's surrogate modeling system. “And if you solve a full model with unknown inputs, it might take you 15 minutes. And people are very impatient.”

According to Sjödin, European car manufacturers are now using COMSOL surrogate models to quickly simulate entire processes. electric car battery packs enabling real-time decisions for which managers and engineers once had to wait for a coffee break or longer. Meanwhile, Sjodin adds, the Swiss institute has implemented the COMSOL surrogate system as an application for Indian farmers to predict food spoilage during cold storage. The institute found that surrogate modeling allowed farmers to reduce food waste by 20 percent.

A full COMSOL numerical simulation predicts the antenna's surface performance (right sphere), while its optimized surrogate model (left sphere) produces almost the same results in significantly less runtime.KOMSOL

Turning multiphysics into an application

Sjödin says COMSOL intends to transform users of the simulation system into something closer to software developers in my own way.

“You can compile these applications into individual executable files that can be distributed worldwide without any licensing fees,” says Sjodin.

The company's surrogate models can operate as standalone applications that can run on laptops or smartphones.

“If you want to transfer this to someone in a plant, these surrogate models are really useful because they allow you to immediately evaluate and get results,” says Sjodin. Models work faster compared to full-fledged ones. multiphysics modeling because a version of the app about, say, the thermal characteristics and chemical composition of a particular battery pack is preloaded. Simulations are fast because there are already pre-calculated parameters specific to the physical environment being simulated, and only to the environment being simulated.

In addition to AI intelligence that speeds up computation time every time you run it, COMSOL uses other tricks as well. What do modelers call “reduced order” models (ROM) include optimizations such as mathematical pattern recognition and shortening some of the more complex equations in calculations. “Neural networks This is where other technologies, more traditional reduced-order modeling technologies come into play as well,” he says.

For example, in 2024 Industry-Wide Review of ROMsresearchers from Trieste, Italy International School of Advanced Studies described a number of ROM methods that are based not only on artificial intelligence or neural networks.

“ROMs fall into two broad families: intrusive methods, in which the governing equations can be directly manipulated, and non-intrusive methods, in which only simulation data are considered,” the researchers wrote. The paper shows that the combination of neural networks and more traditional ROM math tools can achieve computational acceleration up to 100,000 times faster than models without added smart ROMs.