EP-4742078-A1 - REAL-TIME SIMULATION OF INDUSTRIAL PARTICLE FLOWS

Abstract

A method for real-time simulation of material flow in an industrial process involves receiving input data that characterizes the physics parameters of the material flow. This input data is used to generate a continuous field representation, which feeds into a trained machine-learning model. The model then predicts the material flow, and this prediction result is displayed on an electronic display.

Inventors

• Brandstetter, Johannes
• Kronlachner, Tobias
• LICHTENEGGER, THOMAS
• Alkin, Benedikt

Assignees

• NXAI GmbH

Dates

Publication Date

20260513

Application Date

20241112

Claims (15)

1. A method of real-time simulation of a material flow in an industrial process, comprising: receiving input data that characterizes one or more physics parameters of the material flow in the industrial process; generating a continuous field representation of the input data; generating, using a trained machine-learning model that takes as input the continuous field representation, a prediction of the material flow in the industrial process; and displaying a result of the prediction on an electronic display.
2. The method of claim 1, wherein the material flow comprises one or more of: a particle flow; a fluid flow; a coupled particle-fluid flow.
3. The method of claim 1 or 2, wherein the step of displaying a result of the prediction is performed while the simulation is running and/or while the industrial process is running.
4. The method of any one of claims 1 to 3, wherein the step of generating a prediction of the material flow is performed while the simulation is running and/or while the industrial process is running.
5. The method of any one of claims 1 to 4, wherein the continuous field representation is configured to model a Lagrangian discretization of a Discrete Element Method (DEM) routine in a compressed latent space.
6. The method of any one of claims 1 to 4, wherein the machine-learning model comprises a neural operator.
7. The method of claim 6, wherein the neural operator is configured to calculate a numerical Discrete Element Method (DEM) routine.
8. The method of claim 6 or 7, wherein the neural operator is a multi-branch neural operator comprising at least one main branch and at least one auxiliary branch; wherein the at least one main branch is configured to model a primary quantity that represents one or more microscopic properties of the material flow and/or the industrial process, preferably as continuous field representation; wherein the at least one auxiliary branch is configured to model a secondary quantity that represents one or more macroscopic properties of the material flow and/or the industrial process, preferably as continuous field representation.
9. The method of claim 8, wherein the multi-branch neural operator comprises a plurality of main branches, wherein the corresponding primary quantities influence each other.
10. The method of claim 8 or 9, wherein the multi-branch neural operator comprises a plurality of auxiliary branches, wherein the corresponding secondary quantities are independent.
11. The method of any one of claims 8 to 10, wherein each main branch and auxiliary branch comprises a stack of transformer blocks where weights are not shared between branches.
12. A trained machine-learning model configured for use in the method of any one of claims 1 to 11.
13. A data processing apparatus comprising means for carrying out the method of any one of claims 1 to 11.
14. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any one of claims 1 to 11.
15. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of any one of claims 1 to 11.

Description

TECHNICAL FIELD The present invention generally relates to the field of neural network system architectures for machine learning, and more particularly to hardware-efficient neural network models which are usable for physics simulations. BACKGROUND Advancements in computing power have made it possible to numerically simulate large-scale fluid-mechanical, particulate, or otherwise material flow-based processes. Such simulations are increasingly important for understanding and optimizing industrial processes in various sectors, including pharmaceuticals, agriculture, chemical engineering, and engineering in general, such as relating to granular flows and powder mechanics. The Discrete Element Method (DEM), see Cundall, P. A. and Strack, O. D. L. A discrete numerical model for granular assemblies. Geotechnique, 29(1):47-65, 1979, provides one of the most popular representations of a wide range of physical systems involving particulate matter. DEM is particularly well-suited for capturing the interactions of individual particles, providing insights into the bulk properties of the material, by tracking and computing the behavior of each grain or particle. Consequently, DEM has become a widely accepted approach for tackling engineering problems connected to disperse and discontinuous materials, particularly in granular flows and powder mechanics. Typical target areas comprise mining and mineral processing, steelmaking, pharmaceutics, agriculture and food processing, and additive manufacturing and powder bed fusion. However, DEM simulations are computationally expensive because of the intrinsic multiscale nature of particulate systems, restricting either the duration of simulations or the number of particles that can be simulated. This limits the practical applicability of DEM for certain industrial problems. Moreprecisely, the inherent multiscale nature of particulate systems makes DEM computationally infeasible in several regards: 1. Large-scale granular flows typically consist of a huge number of particles with each of them interacting with the surrounding ones. For every grain, its equation of motion (EOM) needs to be solved in a coupled fashion. Current DEM studies dealing with process relevant problems use up to two million particles, while e.g. an industrial shaft furnace or fluidized bed reactor may well contain several orders of magnitude more grains.2. The high material stiffness of solid particles severely limits the numerical time step that can be used in the solution procedure of the EOMs. Often, its value is in the range of microseconds, whereas process-relevant durations may be minutes or hours.3. There is no straight-forward relationship between the microscopic DEM parameters and macroscopically observed behavior. Instead, optimization techniques need to be used to find a set of DEM parameters that reproduces certain characterization measurements (e.g. angle of repose and shear cell measurements). Such a calibration routine needs to be performed for any kind of material before the actual simulation of interested can be approached. Any change of material properties (e.g. due to different size distribution, degradation, moisture uptake) requires a new calibration. The first-mentioned issue above is usually mitigated by employing coarse-graining techniques that replace many small particles with a large parcel. If the interaction parameters of these parcels are chosen appropriately, either using scaling rules or a calibration routine, the accuracy impairments compared to the fine-grained ground truth are often acceptable. However, the limitation of small time steps and the need for parameter calibration persist and make DEM slow and sometimes too cumbersome for a quick application within engineering workflows. Furthermore, establishing a reliable connection between the microscopic parameters used in DEM simulations and the macroscopic material properties observed in experiments often requires elaborate calibration procedures. This calibration process can be time-consuming and may involve significant uncertainties. It is therefore an objective of the present invention to provide improved methods and systems for simulating industrial material flows that require less computing resources, thereby overcoming the above limitations at least in part. SUMMARY OF THE INVENTION The above and other objectives may be achieved by the subject-matter defined by the independent claims. Advantageous modifications of embodiments of the present disclosure are defined in the dependent claims as well as in the description and the drawings. One aspect of the invention relates to a method of simulation. The method may be a method of physics simulation. The method may be a method of simulation of a technical system, device and/or process. The method may be a method of simulation of a material flow. The material flow may be in an industrial process. The method may be performed within an engineering workflow. The simulation may