romAI: How to speed-up DEM simulations leveraging Artificial Intelligence
Altair EmployeeAbstract
In this article we will use a novel AI-based application, romAI, to generate a dynamic Reduced-Order-Model (ROM) from simulation Data. We will demonstrate, on a system simulation example, how this innovative approach can support the design of heavy machinery. romAI will be coupled with a wheel loader model to estimate the loads due to the bulk material interaction, significantly reducing the simulation runtime around 35 times.
Introduction
On this application we discuss the analysis of a wheel loader used to handle bulk material. This system can be simulated with the aid of Altair’s software. The Wheel Loader is represented in the Multibody software MotionSolve® and includes Flexible bodies created in OptiStruct®. PID controllers are used in the boom and bucket actuators to achieve the desired position. Finally, the bulk material is simulated with EDEM® which is a Discrete-Element-Method (DEM) solver.
As expected, different handling maneuvers vary the loads from the bulk material to the bucket and thus the loads on the individual parts, e.g., hydraulic actuator forces and stresses on critical components.
Traditionally for each different maneuver, we would have to run a co-simulation with EDEM. Nevertheless, DEM simulations in general can be time costly, particularly on large scale problems.
The Challenge
One of the objectives for these analyses is to test different maneuvers and estimate stress and loads on the critical components with good accuracy.
Can Artificial Intelligence (AI) help us to speed up the study by replacing the DEM system with a more time efficient but still accurate Reduced-Order Model ??
Workflow
We’ve followed a 3-step workflow:
- Data Generation
- ROM generation & Validation
- ROM deployment / Results evaluation
It took less than 9 hours to accomplish them all !
To generate the required data for training the romAI we have run 5 different scenarios with DEM system only performed in EDEM. In each scenario the “X, Z” displacements and “θ” bucket angle, along with their respective velocities were varied slightly, to cover an adequate range of motion.
We have then saved the data into a csv file and fed it to the romAI GUI for the generation of the dynamic reduced-order model.
The inputs to the system are the already mentioned displacements, angles and velocities while the outputs are the forces along “X” and “Z” directions plus the torque around the “Y” axis. The mass of the material on the bucket represents the state of the dynamic system and also an additional output for us.
Next, we validate the romAI model by comparing its outputs with the EDEM ones for the same inputs.
Once we have validated the generated ROM, we connect the romAI with the wheel loader model and run a Co-Simulation inside Activate.
Finally, to have a reference for the comparison, we run the co-simulation with EDEM on the same maneuver.
In such loading scenarios it is important to estimate the required actuator forces in order to properly size them. Below are the results of the hydraulic forces to perform these movements. The difference on the max force is bellow 1.4 % !
One of the most critical components on the Wheel Loader is the Boom, as it is responsible for transmitting the loads from the bucket to the frame and the hydraulic actuators. Using the Flex body modal analysis approach in MotionSolve we are able to calculate the Stresses on the Boom, which may be used to calculate the lifecycle of the part under fatigue.
It is remarkable that the difference on the max Von Misses stress on the critical element between the original model and the romAI equipped one is under 2% !!
Conclusion
- A proper reduced-order model was generated starting from data obtained by few EDEM simulations, covering an adequate range of motion.
- The high non-linearity of the DEM model makes this approach even more challenging. Nevertheless, romAI allowed to predict the stress and loads on critical components with a small deviation.
- The romAI can estimate the bulk material forces even on maneuvers that deviate from those used on the training while reducing the simulation runtime around 35 times: from 680s to 20s !!
Do not hesitate to contact Altair for additional information on System Simulation combined with AI !