The research shows that deep learning models can realistically reproduce physical processes over long periods of time. The researchers are also able to generalize across different simulation parameters and geometries. The team is aiming for fast real-time simulations, is planning to set up foundation models for industrial customers and will focus on the generalization of simulations in the next step. He will explain how this will work at our Industrial AI conference in January (LINK)

Discrete Element Methods (DEMs) are the industrial standard in the simulation of granular flows and powder simulations. In addition, the numerical calculation method of particles also plays a major role in the simulation of chemical processes. However, DEMs have disadvantages: they are computationally intensive and often complex to calibrate. The NXAI and JKU Linz research team led by Brandstetter combine neural networks and DEMs in their NeuralDEM model and promise faster simulations, parameter optimization and industrial simulations in real time.

The basis for the latest research successes is the NXAI architecture of Universal Physics Transformers (UPT). This is a method of improving neural networks so that they can process very large amounts of data faster, more efficiently and learn physics in an abstract, compressed representation of the physical world. “Encoding information and then decoding it is the challenge,” explains Benedikt Alkin from the NXAI research team. UPTs are not dependent on traditional approaches such as lattice structures (for example in flow simulations) or particle models. “UPTs can be applied to a wide variety of simulations,” adds NXAI colleague Tobias Kronlachner.

“AI and neural networks in particular have now arrived in the world of simulation. Thanks to UPT, our neural network learns the physics and we prove that our AI-based simulations reliably learn the physics and then reproduce it. This is the decisive criterion for acceptance in industrial applications,” emphasizes Brandstetter.

The NXAI team has been working on the model for over a year. With Prof. Dr. Stefan Pirker from the JKU Linz, Thomas Lichtenegger and Tobias Kronlachner, the Linz team included three domain experts. Samuele Papa from the AI powerhouse in Amsterdam provided support in deep learning. The researchers demonstrated the performance of the model in various transport processes, including mass, species and residence time. They visualize this in three scenarios: the emptying and refilling of silos with varying discharge angles and fluidized beds with different inflow velocities. The model generated true-to-life physical simulations. The largest NeuralDEM model is capable of faithfully modeling coupled CFD-DEM fluidized bed reactors with 160k CFD cells and 500k DEM particles for trajectories of 28s, i.e. 2800 ML time steps.

Brandstetter and his team will reveal what they want to do with the results, which industries they are focusing on and what other ideas the Austrians have at our Industrial AI conference in Frankfurt on January 22.