The NXAI research team led by Ass. Prof. Dr. Johannes Brandstetter is the first team worldwide to present NeuralDEM, an end-to-end deep learning alternative for modifying industrial processes such as fluidized bed reactors or silos. The research shows that deep learning models can realistically map physical processes over long periods of time. In addition, the researchers are able to generalize across different simulation parameters and geometries. The team is aiming for fast real-time simulations, plans to build foundation models for industrial customers and is focusing on the generalization of simulations in the next step.

Discrete Element Methods (DEMs) are the industry standard in the simulation of granular flows and powder simulation. In addition, the numerical calculation method of particles also plays an important 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 Johannes Brandstetter are combining neural networks and DEMs in their NeuralDEM model, promising faster simulations, parameter optimization and industrial simulations in real time.

A neural network learns physics

The basis for the latest research successes is the NXAI-patented 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 and more efficiently and learn the physics in an abstract, compressed representation of the physical world.

“AI and especially neural networks have now entered the simulation world. Thanks to UPT, our neural network learns the physics and we prove that our AI-based simulations reliably then map it again. This breaks existing barriers for industrial applications,” emphasizes Johannes Brandstetter.

The NXAI team has been working on the model less than a year. The Linz-based team brought three domain experts into the team: Prof. Dr. Stefan Pirker from JKU Linz, Thomas Lichtenegger and Tobias Kronlachner. Samuele Papa from the AI powerhouse Amsterdam provided support in deep learning. The researchers demonstrate the model's performance in various transport processes, including mass, species, and residence time. They visualize this in three scenarios: the emptying and refilling of silos with varying outflow angles, and fluidized beds with different incident flow velocities. The model produced accurate 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.

Further Information: For more details, visit our project page.

Brandstetter sees NXAI as a pioneer in industry. “On the one hand, we have proven that we are able to develop very large models and calculate them efficiently. On the other hand, we have extensive domain knowledge in the field of simulation, which sets us apart from others. Those who can offer similar computing power often don't have this specific expertise and usually show little interest in acquiring it. At the same time, the simulation field experts often lack experience with large neural networks. This is precisely where our unique selling point lies: we combine expertise in large networks with in-depth domain knowledge.” In the near future, NXAI aims to build simulation foundation models, based on other types of simulations such as CFD (computational fluid dynamics), and offer these to customers. “We have an amazing team of domain and machine learning experts and NeuralDEM is only the beginning,” promises Brandstetter.