HEEDS/ DARS-Basic Global Mechanism Optimization Megan Karalus, PhD Application Engineer CD-adapco November 2014
Why do I need a global mechanism? Simple Chemistry Global Mechanism STAR-CCM+ Predict CO Emissions Flame Behavior
What is a global mechanism? Level of Description Reactions Notes Single Step CH 4 + 2O 2 CO 2 + 2H 2 O Complete Combustion Three Step Detailed Kinetic Mechanism Take Methane as an Example CH 4 + 1.5O 2 CO + 2H 2 O CO + 0.5O 2 CO 2 CO 2 CO + 0.5O 2 Hundreds Species: OH, O, H, CH, CH 2 O, C 2 H 6, etc Includes Some Intermediate Species Rates are fitted Valid for a narrow range of conditions Includes all intermediate species Rates are measured Valid for a wide range of conditions What are the rates of the global mechanism????
Process Surrogate Model Analysis Software Optimization Software/ Algorithm
Surrogate Model and Analysis Software Compare against results using detailed mechanism Can t run detailed in CFD. Need Rates for Global Mechanism Need Surrogate Model Allows us to focus on kinetics Can handle detailed mechanism to generate target values Freely propagating laminar flame. DARS-Basic
DARS-Basic Simulation Freely propagating flame (fuel and air are premixed) Fuel/Air Premix Hot Products Reaction Mechanism describes this part
Process Surrogate Model Optimization Software/ Algorithm
HEEDS MDO HEEDS MDO is a multi-disciplinary optimization tool from Red Cedar Technology. There are two components to HEEDS MDO: Process Automation Automate the Virtual Prototype Build Process Enable Scalable Computation across platforms Design Exploration Efficient Exploration (Optimization, Sweeps, DOE) Sensitivity & Robustness Assessment Process Automation Design Exploration These two components combined, coupled with its leading hybrid adaptive search algorithm SHERPA, makes HEEDS MDO the most technologically advanced parametric optimization tool in the world
9 Standard Optimization Process Build Baseline Model Define Optimization Problem Select Optimization Algorithm and Set Tuning Parameters Proposed Solution No Satisfied? Yes Optimized Solution
10 Standard Optimization Process Build Baseline Model Define Optimization Problem Characteristics of the design space are unknown Select Optimization Algorithm and Set Tuning Parameters Proposed Solution No Satisfied? Yes Optimized Solution
11 Standard Optimization Process No Build Baseline Model Define Optimization Problem Select Optimization Algorithm and Set Tuning Parameters Proposed Solution Satisfied? Yes Gradient-based methods Linear programming Simplex methods Genetic algorithm Simulated annealing Particle swarm method Ant colony method Response surface methods Etc. Optimized Solution
12 Standard Optimization Process No Build Baseline Model Define Optimization Problem Select Optimization Algorithm and Set Tuning Parameters Proposed Solution Satisfied? Yes Genetic algorithm (GA) Population size Number of generations Cross-over type Mutation type Selection type Cross-over rate Mutation rate Selection parameters Etc. Optimized Solution
13 Modern Optimization Process Standard Procedure Build Baseline Model HEEDS Procedure Build Baseline Model Define Optimization Problem Define Optimization Problem Select Optimization Algorithm and Set Tuning Parameters Proposed Solution SHERPA Hybrid, Adaptive Optimization Algorithm No Satisfied? Yes Optimized Solution No Tuning Parameters No Opt Expertise Required Optimized Solution
SHERPA Search Algorithm The SHERPA Search Algorithm Hybrid Blend of search strategies used simultaneously Global and local search performed together Leverages the best of all methods Adaptive Adapts itself to the design space Efficiently searches simple and very complicated spaces Very cost effective for complex problems! 14
Process Surrogate Model Now we look at our study.
Global Mechanism 1) JetA + 2O 2 -> 4C 2 H 4 + 4CO + 3.5H 2 2) C 2 H 4 + O 2 -> 2CO + 2H 2 3) CO+ H 2 O = CO 2 + H 2 4) CO 2 -> CO + 0.5 H 2 O Pressure = 3.5 bar, Temperature = 450K, Equivalence Ratio = 0.4 4.8 5) H 2 + 0.5O 2 -> H 2 O Honeywell F. Xu, V. Nori, J. Basani. CO Prediction for Aircraft Gas Turbine Combustors. Proceedings of the ASME Gas Turbo Expo 2013. GT2013-94282.
What do we need for Optimization Study? Variables What are they? Range to vary? Initial guess (Baseline) Responses How do we evaluate results? Objectives How do we measure improvement? Constraints Do we need to constrain?
10 Variables Sample Reaction: Reaction Rate: C + D E + F ω = Ae E A/RT C n D m Variables we can vary A : Pre-exponential Factor n, m : FORD (forward reaction rate exponents)
4 Responses (Curve Fits) CO vs. T (K) Phi = 0.6 CO vs. T (K) Phi = 1.0 CO vs. T (K) Phi = 1.4 Flame Speed (cm/s) vs. Phi Blue Red = Target (Dagaut) = Baseline
Objectives and Constraints Objectives Weight Curve Fit: Flame Speed 1 Curve Fit: CO vs. T, Phi = 0.6 20 Curve Fit: CO vs. T, Phi = 1.0 10 Curve Fit: CO vs. T, Phi = 1.4 10 Constraints Flame Speed Error at Phi=1.0 +/- 10% Max CO Error at Phi = 0.6 +/- 10%
SHERPA Benchmark Example HEEDS Optimization Design Curve Target Curve Change design variables SHERPA Responses Note that only the CO (0.6 value for phi) objective history plot is shown 21 Design Exploration
SHERPA Benchmark Example HEEDS Optimization Design Curve Target Curve OPTIMIZED DESIGN Change design variables SHERPA Responses Note that only the CO (0.6 value for phi) objective history plot is shown 22 Design Exploration
HEEDS Results 1000 Evaluations 5.5 hours
Results: Parallel Plots
Results: Parallel Plots
Percent Change from Baseline 1) JetA + 2O 2 -> 4C 2 H 4 + 4CO + 3.5H 2 2) C 2 H 4 + O 2 -> 2CO + 2H 2 3) CO+ H 2 O = CO 2 + H 2 4) CO 2 -> CO + 0.5 H 2 O 5) H 2 + 0.5O 2 -> H 2 O A_1 1E+12 3.41E+11-66% A_2 1E+12 1E+12 0% A_3 5E+12 7.49E+12 50% A_4 2.00E-08 5.12E-08 156% A_5 1E+14 1.01E+14 1% _1_FORD_JetA 0.5 0.544 9% _1_FORD_O2 0.6 0.59-2% _2_FORD_C2H4 0.8 0.788-2% _2_FORD_O2 0.8 0.82 2% _4_FORD_H2 0.5 0.552 10% _4_FORD_O2 1.2 1.2 0%
HEEDS Results CO vs. T (K) Phi = 0.6 CO vs. T (K) Phi = 1.0 CO vs. T (K) Phi = 1.4 Flame Speed (cm/s) vs. Phi Blue Red = Target (Dagaut) = Baseline Green Purple = Best Design = Honeywell
Testing this mechanism across full range.
Summary of Multi-Objective Study Large range explored for each variable (non-error designs) Many feasible designs found. Best designs are mostly clustered around same solution. Focusing on smaller equivalence ratio range sped overall computations with little cost to the final optimized mechanism. Adequately captured best result from manually tuned global mechanism in ASME paper.
Can we do better? Multi-objective optimization showed significant improvements over baseline. We know (from experience and the paper) that there is a trade-off in predicting CO vs. Flame Speed. Specifying how much we re willing to compromise on one or another can be difficult -> Trade-off Study to find Pareto Front. Trade-off Study also helps illuminate the underlying limitation of the global mechanism chosen for optimization. Competing Objectives: Flame Speed Curve Fit Unified CO Curve Fit
Pareto: Trade-off Study Best Compromise
Pareto: Trade-off Study Better CO
Pareto: Trade-off Study Better Flame Speed
Pareto Front Conclusions Confirms trade-off in predicting Flame Speed and CO. Provides additional information on limitations of chosen global mechanism. Gives engineer multiple options, depending on goal of CFD simulation.
Thank you! Questions?
HEEDS Software Efficient Exploration Benchmark Function : f x n å ( ) ( ) = - x i sin x i i=1-500 x i 500 Minimum : f = -418.9829n Graph showing function for 2 variables x 1 x 2 Results for n = 20 Average values for 25 optimizations from random baselines 36 Copyright 2014 - Red Cedar Technology: All Rights Reserved