(40b) Impact of Computational Mesh on CFD Combustion Predictions

Use of computational fluid dynamics (CFD) for combustion modeling has become widespread throughout the petrochemical industry. Commercial CFD software such as CD-Adapco’s StarCCM+ and ANSYS’s Fluent are routinely used by designers, vendors, consultants and end users to evaluate burner and furnace design and performance. For example, CFD models have been used to predict flame length, flame impingement, heat flux profiles, heat transfer, and outlet gas properties. Because combustion is a complex interaction of turbulent mixing, chemical reactions, and heat transfer, development of accurate CFD combustion models requires familiarity with both combustion principles and modeling options in the CFD software. Specifically, modelers are required to select appropriate computational grids, boundary conditions, turbulent flow and chemical reaction sub-models, and numerical solution techniques. The specification of grids and boundary conditions and selection of appropriate sub-models is complicated by the numerous options available in commercial CFD codes. These options are offered to aid modelers in accurately representing the wide variety of physical conditions that occur in different combustion processes, but to inexperienced users the preponderance of gridding and sub-model options can be confusing, particularly when specification of an incorrect grid or boundary condition or selection of one incorrect sub-model can impact the accuracy of the overall prediction. These risks of incorrect model setup or sub-model selection can be mitigated as users gain additional training, experience, and exposure to CFD modeling.

This paper provides information related to one aspect of combustion modeling setup - generation of the computational grid. Grid generation encompasses the type, quantity and localized refinement of computational cells used to represent the geometry and interior volume of the furnace. Inexperienced CFD users are often not familiar with the impacts of selecting different types or refinement levels of computational grids. This paper presents results of combustion simulations of a fuel-staged burner in a test furnace using the commercial CFD code StarCCM+. Simulations have been completed using different grid types, quantity of grid cells, and local grid refinement to determine the impact of these factors on predicted flame shape, flue gas properties, incident heat flux profile, and process tube heat transfer.

The application selected for study was representative of a Vertical Cylindrical (VC) heater burner in a test furnace. The test furnace was 10.5 ft wide by 12 ft deep by 41.5 ft tall, with four 6.625-in diameter vertical (floor to roof) cooling tubes spaced evenly along the back wall; tube centerlines were 1.5 ft from the back wall. The furnace exit was a 3 ft by 3 ft opening centered over the burner. For simplicity, all furnace walls were assumed to be adiabatic. Cooling tubes were assigned a fixed temperature of 100 °F and emissivity of 0.85. A 10 MBtu/hr burner was centered in the box. The burner was representative of low NOx burners used in VC heaters, with ten staged firing tips arranged symmetrically outside a sloped burner tile at an effective diameter of 28.5 in. Air flowed upwards in the center of the burner through a 16.75-in diameter throat. Fuel was 100% methane at 60 °F distributed evenly through the ten fuel tips. Inlet air temperature was 60 °F. The burner was operated at a stoichiometry of 1.15, with 3% excess O₂ (dry). There was no premixing of fuel and air in the burner.

Simulation results for five computational meshes or grids will be summarized in the paper. All cases had the same burner, cooling tube and furnace geometry, the same boundary conditions, and the same fuel and air properties. The following standard StarCCM+ sub-models were used for all cases: k-e turbulence sub-model, Presumed Probability Density Function (PPDF) chemical reaction sub-model with equilibrium chemistry, discrete ordinates radiation solver with weighted gray gas sub-model for participating medium. The computational grids for each case are summarized as follows:

Case 1 - a 1-million node, mixing zone refined, trimmed cell approach. The StarCCM+ trimmed cell gridding option was used to create a structured-like grid comprised of 982,263 cells. The grid was refined around the fuel tips, burner and process tubes to accurately represent the geometrical shapes and regions of greatest fuel-air mixing around and above the burner.

Case 2 - a 2-million node, mixing zone refined, trimmed cell approach. The StarCCM+ trimmed cell gridding option was used to create a structured-like grid comprised of 2,097,728 cells. The grid had an additional level of refinement around the fuel tips, burner and process tubes relative to Case 1.

Case 3 - a 1-million node, inlet zone refined, trimmed cell approach. The StarCCM+ trimmed cell gridding option was used to create a structured-like grid comprised of ~1 million cells. The grid was refined at the fuel tip inlets and air inlet, but not above the burner in the fuel-air mixing zone. Sufficient refinement was used to accurately represent the shape of the cooling tubes.

Case 4 - a 2-million node, mixing zone refined, polyhedral cell approach. The StarCCM+ polyhedral cell gridding option was used to create a polyhedral grid comprised of 1,970,394 cells. The grid was refined around the fuel tips, burner and process tubes to better represent the geometrical shapes and regions of greatest fuel-air mixing. StarCCM+ default gridding procedure for this option caused the grid refinement to be somewhat biased towards the cooling tubes.

Case 5 - a 1-million node, mixing zone refined, polyhedral cell approach. The StarCCM+ polyhedral cell gridding option was used to create a polyhedral grid comprised of ~1 million cells. The grid was less refined around the fuel tips, burner and process tubes relative to Case 4. This simulation is not completed.

The five cases were chosen in order to compare the effects of number of grid cells (e.g., 1 million vs 2 million cells), grid type (e.g., trimmed cell vs polyhedral), and grid refinement (mixing zone refinement vs inlet zone refinement). Results were compared based on computational time (time per iteration), computer memory used, flame shape based on a 1500 ppm iso-surface of CO concentration, furnace exit gas temperature (FEGT) and O₂ concentration, total heat transfer to the cooling tubes, and the radiative incident heat flux profile to the cooling tubes. The following key results were observed:

1) There were minimal differences in flame shape between the 2-million node trimmed cell and polyhedral cell cases (Case 2 and Case 4). Both showed similar flame length and width with minor CO cusps near the top of the burner where the greatest fuel and air mixing occurred. Heat flux profiles were also similar with peak fluxes near 100 kW/m² occurring at approximately 60% of the furnace height.

2) Reducing the trimmed cell mesh from 2 million to 1 million cells (Case 2 vs Case 1) caused minor differences in flame shape. The flame was slightly broader and the CO cusps near the top of the burner were slightly more pronounced, but the overall flame shape was similar. Heat flux profiles also reflected these minor differences, with Case 1 predicting 2-3% higher heat fluxes.

3) When the trimmed cell mesh was refined in the inlet zone rather than the high fuel-air mixing zone above the burner (Case 3 vs Case 1), the flame shape was notably different. The Case 3 flame was predicted to be much longer, with higher CO concentrations in the upper furnace. This was a result of the fuel jets not penetrating and mixing with the burner air. The relatively coarser mesh in the mixing zone directly above the burner limited the fuel and air mixing and resulted in CO and oxygen moving adjacent to each other upwards with the flame. This also resulted in a shift in the heat flux profile. This grid change caused the most pronounced differences seen between any of the cases.

4) Predicted exit flue gas temperatures were within 1.3% for Cases 1, 2 and 4. Case 3 temperatures were lower due to the incomplete reactions between the CO and oxygen in the furnace.

5) Different grids impacted calculational times. On the four-processor desktop computer used for simulations, normalized calculational iterations per minute were: Case 1 - 1.0, Case 2 - 0.380, Case 3 - 1.0, Case 4 - 0.254, Case 5 - TBD. An estimate of average computer memory usage showed normalized values of approximately: Case 1 - 1, Case 2 - 6.5, Case 3 - 11.

Current results suggest that the 1-million node trimmed cell grid (Case 1) is a reasonable option for this simulation based on similar flame shape, heat flux and gas property results coupled with faster runtimes and lower memory requirements (relative to Cases 2 and 3). However, the similarly-sized grid used in Case 2 is not recommended because the local grid refinement did not accurately represent the fuel-air mixing just above the burner. Grid refinement appeared to cause the most significant differences between the cases studied. The results should serve as useful guidelines to combustion modelers specifying computational meshes for similar combustion applications.