Using Machine Learning and Simulation to Optimize a Rotary Dryer
Using Machine Learning and Simulation to Optimize a Rotary Dryer
1. Introduction
Aggregate drying is a fundamental unit operation in asphalt manufacturing but is energy intensive and requires a great degree of control to reliably achieve target product properties in a wide range of ambient conditions and using diverse feedstock.
This example involves the implementation of the recorded factory feature. The recorded factory folder, 192_start_withvolumeadded_4coupling_2023.1_New Section 10_recording, provided has been simplified to reduce the file size. Recorded Factories allows you to record the material flow of a simulation, and run multiple scenarios, by reusing the recorded mass flow of your granular simulation. For further information, please refer to:
All the requisite files can be downloaded below:
A common drier design is shown in Figure 1 and uses hot air to dry aggregate particles flowing through a rotating drum with lifters.
Figure 1: Aggregate rotary dryer
The thermo-mechanical behaviour of this system is not fully understood, and equipment design is largely empirical. Consequently, poor energy efficiency and sub-optimal reliability are common issues. Producing improved designs is key to increasing profitability in manufacturing operations and reducing greenhouse gas emissions but requires data which is time consuming and expensive to obtain via physical trials alone. High-fidelity physics-based virtual prototyping and optimization can provide a rapid and cost-effective alternative in this context.
This article presents an efficient virtual optimization methodology that integrates CAD modelling, geometry parametrization and high-fidelity physics-based simulation in Altair® Inspire™, Altair® SimLab™, Altair® EDEM™ and Altair AcuSolve™ as well as machine learning optimization and automation in Altair® HyperStudy® to rapidly identify optimal designs and operating conditions for aggregate drum driers. The workflow aims to optimize the internal lifter geometry (expanded in Figure 5) in an aggregate dryer to increase its drying efficiency.
2. Workflow
2.1 Overview
A rotary dryer is modelled using a coupled CFD-DEM simulation approach. The asphalt aggregate is modelled in Altair EDEM, a high-fidelity bulk material simulation software powered by Discrete Element Method technology. The gas phase is modelled in Altair SimLab, which is one of the multi-physics pre-processors available in the Altair portfolio and is used here as the interface for Altair’s CFD solver AcuSolve.
As the dryer rotates, the lifters produce a veiling which is affected by the lifter angle as shown by the vector representation of the particles in Figure 2. A good balance of cascading and particle mass in the flow stream of hot air generated by a gas burner, allows for greater contact between the fluid and particles leading to effective drying. The aim of the optimisation is to identify an optimal lifter angle for a given set of gas flow properties.
Figure 2: Particle veiling affected by lifter angle. Particles shown by vector representation
The optimisation methodology is summarized in Figure 3.
Figure 3: Optimisation workflow overview
The CAD geometry of the drier is generated and parametrized in Altair Inspire; The Inspire database is imported into SimLab utilising the SimLab-Inspire Extension, shown in an example here, allowing the CAD variables to be defined within SimLab and to be optimised later. A coupled Flow + EDEM solution is created, parameterized and recorded as a python macro in SimLab.
Altair EDEM is used to model the particle phase by defining the bulk material, particle factory, particle physics models and geometry kinematics. Finally, the SimLab python script and the EDEM simulation deck are imported into Altair HyperStudy as models, where the case generation is automated, and the design optimization is performed.
2.2 Geometry parametrization using Altair Inspire
The optimization of the aggregate drum drier design begins with geometry parameterization, which is conducted in Altair Inspire. The design parameter space is defined by assigning variables to the sketch and geometry tools, and the Variable Manager within Inspire is used to update these parameter values to alter the design of the drier. The CAD parameter space is summarized in Figure 4 and focuses on the lifter geometry, which affects the cascading of the particles and, therefore, the drying rate. For further details on geometry parametrization, please review – Inspire Variables.
Figure 4: Geometry parameter space
Figure 5 shows the different lifter design configurations produced after the variables have been altered.
Figure 5: Different lifter configurations
2.3 Multi-Physics Modelling using Altair SimLab and Altair EDEM
The system is reduced to the section of four lifter rows shown in Figure 6 in the interest of computational efficiency.
Figure 6: Reduced section (Blue) from full geometry
The boundary particle flow is transferred from the full-scale simulation of the drying process using the Altair EDEM recorded factory feature. Using this feature, all particles crossing a virtual plane in the full-scale model are recorded and reproduced in the reduced model as shown in Figure 7. The flow of particles is recorded at steady state over a single drum revolution and continuously looped in the reduced model at the same frequency.
Figure 7: Recorded factory implementation
Coupled simulations using Altair AcuSolve and Altair EDEM are used to simulate the drying process with varying lifter designs. The initial CAD model is created in Inspire and imported into SimLab to parametrize the geometry variables using the SimLab variable manager. The updated lifter geometry CAD model is exported from SimLab as an .stl file to later be imported into Altair EDEM and replace the existing EDEM geometry.
A python macro script is recorded in SimLab to automate the model set-up and parametrization, shown in an example here.
2.4 DoE Generation and Machine Learning Optimisation using Altair HyperStudy
2.4.1. DoE Generation
The case generation is automated in Altair HyperStudy and consists of two models.
a.) The SimLab model – Imports the Inspire - CAD database, parametrizes the geometry variables for a given design configuration, exports a surface mesh geometry, creates the fluid domain and sets up the CFD case including the coupled flow solution and variables.
b.) The EDEM model – Imports the surface mesh geometry exported by SimLab and couples the EDEM deck with the AcuSolve solver which will execute the SimLab generated input file.
The model configuration is summarized in Figure 8.
Figure 8: HyperStudy model configuration
The EDEM-HyperStudy connector allows the customisation for EDEM variables and accepts solver input arguments to import and replace the geometry using the -stl flag in solver input arguments, shown in Figure 8, along with coupling the EDEM deck to an AcuSolve .inp file using the -co flag in solver input arguments, shown in Figure 8. Adding the SimLab model in HyperStudy imports all the variables which can be adjusted within HyperStudy.
The active parameter space consists of the lifter angle and the flow velocity, shown in Figure 9.
Figure 9: HyperStudy parameter space
In order to analyse the process performance, a geometry bin and a result export is set up in EDEM to output a Results.csv file containing the total evaporation rate, average particle temperature, total particle heat flux and the total particle volume added (moisture content).
A useful technique to sample the parameter space for statistical analysis or model fitting is to use the HyperStudy Design of Experiments - DoE feature. In this instance, the Fractional Factorial method was used to generate four cases evaluating the extremes of our parameter space and combining that with a second well distributed DoE using the Modified Extensible Lattice Sequence - MELS DoE. The 11 cases listed in Figure 10 were generated in total.
Figure 10: MELS DoE data
2.4.2. Statistical Analysis, Machine Learning and Optimisation
In order to understand the parameter sensitivity, we use the pareto plots generated by HyperStudy after the simulation data is obtained. These are shown in Figure 11 and demonstrate the importance of the lifter geometry.
Figure 11: HyperStudy pareto plots for evaporation rate, volume added, particle temperature and heat flux
Machine learning is then used to fit a response model to the simulation data in HyperStudy. The Fit Automatically Selected by Training - FAST method predicts the values of the Key Performance Indicators - KPIs based on generalised accuracy metrics such as the R2 value. This saves a substantial amount of computational resources and time by avoiding the need to run further simulations.
In this example, the FAST method is used to fit models for the dryer’s KPI - total evaporation rate. The model residuals are shown in Figure 12 and demonstrate the good predictive accuracy of the model.
Figure 12: HyperStudy residual pot for evaporation rate
The fitted response model is used for rapid optimisation using HyperStudy’s Genetic Algorithm - GA. It efficiently identifies the globally optimal parameters for a given objective/goal, which is shown in Figure 13.
Figure 13: HyperStudy objective for Genetic Algorithm optimisation
HyperStudy operates on the fit model to evaluate thousands of different parameter combinations in seconds and identify the optimum. An initial optimization run with the goal of maximising the evaporation rate at a fixed flow velocity led to the evaluation scatter plot and optimum point shown in Figure 14. This run optimized the lifter angle to maximise drying.
Figure 14: HyperStudy evaluation scatter plot and optimum angle
Alternatively HyperStudy’s Multi-Objective Genetic Algorithm - MOGA can be used to optimize parameter combinations for multiple goals (goals could be complimentary or competing with each other). The multi-objective optimization goal in this instance was to minimize the flow velocity for energy saving purposes and maximise the EDEM evaporation rate as shown in Figure 15.
Figure 15: HyperStudy MOGA objectives
Additionally, a weighted sum approach could be applied to the multi-objective optimization run to converge onto one optimum point. This is achieved by assigning weights to the objectives corresponding with their relative importance as shown in Figure 16.
Figure 16: HyperStudy weighted sum approach objectives with relative weighting
By changing the relative weights of the objectives, we can optimize for the ‘most energy efficient’ configuration (by increasing the weight of ‘minimizing flow velocity’) or for the ‘maximum drying’ configuration (by increasing the weightage of ‘maximising evaporation rate’). The result of the weighted sum approach from Figure 16 is shown by the 3-D scatter plot in Figure 17 with the single optimum point marked.
Figure 17: HyperStudy 3-D scatter plot for weighted objectives
2.4.3. Verification
The Genetic Algorithm - single objective optimization run produces a single optimum point. The resulting optimal lifter angle leads to a 3.8% increase in total evaporation rate as shown in Figure 18 when compared to the baseline configuration.
Figure 18: Baseline configuration vs Optimised configuration
The optimization results are verified using HyperStudy’s verification methods whereby the parameters from an optimum point(s) are tested in a simulation run and a delta graph is provided to visualize the error-margin.
The verification run’s delta plot is shown in Figure 19 below, where the blue graph shows the initial prediction, and the red graph shows the verified result.
Figure 19: HyperStudy optimum point verification
3. Conclusion
The Altair tools portfolio allows rapid and efficient virtual optimisation of complex systems that involve particle-fluid flows, like the aggregate dryer shown above which achieved a 3.8% increase in evaporation rate, visualized in figure 20 and also shown in figure 18, consequently increasing the dryer’s energy efficiency.
Figure 20: EDEM evaporation rate
This is accomplished using a tightly integrated and highly automated no-code methodology that combines high-fidelity physics-based simulation with machine learning and optimisation techniques.
For further information on modelling a rotary dryer, visit: How to model a rotary drier using EDEM and AcuSolve.
There are a number of tools to help you get started with this, for example:
- EDEM-HyperStudy Connector
- EDEM Tutorials
- How To Calibrate Material Model with EDEM and HyperStudy (YouTube)
If you are looking for further information on EDEM or other Altair products we have plenty more on Altair Community: