HFSS with HPC for Large Finite Antenna Array Design - Ansys · HFSS with HPC for Large Finite Antenna Array Design ... Solution: Finite Array Domain ... we can import it into the

© 2011 ANSYS, Inc. August 25, 20111

HFSS with HPC for Large Finite Antenna Array Design

8-23-2011


Introduction• Domain Decomposition Review

• Challenges of Solving Finite Arrays

• Finite Array Tool

Examples:1. Vivaldi array

2. Patch array

Efficiency Enhancements

Conclusion

Overview

Presenter

Presentation Notes

Add part on what HPC and DDM is, and how it has been optimized for arrays. (slide 4)


• Finite Array DDM Expands the DDM Capabilities• DDM was first released in HFSSv12

• Distributes a model’s mesh/solution across several computers distributing the RAM

• Solves a model’s full behavior as if run on a single computer

Domain Decomposition Overview

Generalized DDM

Presenter

Presentation Notes

DDM originally built for arrays (repeating structure). We did general DDM, but now optimized for arrays.


1. RAM limitations?

2. Long meshing and solve times?

3. Working with geometry with thousands of parts?

– Takes manual setup time

– Leads to large file sizes

– Is more intensive on the UI

What Challenges Exist for Large Finite Array Modeling that DDM Could Solve?


• Utilizes Replicated DDM Unit Cell to Address Array Concerns

• Geometry and Mesh copied directly from Unit Cell Model• Unit Cell geometry expanded to finite array through a simple GUI

• Adaptive Meshing Process imported from Unit Cell Simulation– Dramatically reduces the meshing time associated with finite array

analyses.

– Mesh periodicity reinforces array’s periodicity.

Solution: Finite Array Domain Decomposition

Presenter

Presentation Notes

Parent child relationship ensures that all unit cells will have the same mesh.


Advantages of the new finite array tool in V14 of HFSS:

1. Solves much BIGGER arrays on the same hardware

2. Obtains ACCURATE results that match HFSS explicit simulations

3. Enables EFFICIENT simulation of large finite arrays utilizing domain decomposition (DDM)

4. Makes it EASY to transform a master/slave unit cell into a finite array

Finite Array Tool Advantages

256 element Vivaldi with metal thickness, under 32GB of RAM!

Embedded element pattern: array tool vs. explicit

Presenter

Presentation Notes

Put labels on array pictures. Put better pictures in here (unit cell plus array).


• Start off with a Unit Cell Model to create the Unit Cell Mesh• Accounts for infinite array behavior

• Unit Cell Simulation is Fast

• Unit Cell Simulation is Memory Efficient

• May use radiation boundary or PML capped unit cell

How It Works!!!

Master / Slave Boundary Pairs Mimic Array’s Periodicity

Radiation Boundary Absorbs Radiated Fields


• Construct the Finite Array• Simple GUI based creation

• Reduces model complexity

• Reduces display issues

• Creates 1 unit cell of additional space around the array edge elements to terminate the array fields in vacuum or infinite ground plane.

How It Works!!!

Defines directions of periodicity based on Master / Slave Boundaries

Defines number of elements in each direction


• Pull the mesh from the unit cell

• Set HFSS so it doesn't perform any additional meshing

How It Works!!!

Presenter

Presentation Notes

Want video of setup

© 2011 ANSYS, Inc. August 25, 201110

• Define a set of computers

• Run the Analysis

• Final results is a full solution to the entire finite array• RAM was distributed across many computers

• Mesh was created efficiently through Unit Cell Simulation

• All Mutual Coupling between elements included

• Edge affect included

How It Works!!!

© 2011 ANSYS, Inc. August 25, 201111

How It Works!!! - Video Demonstration of Finite Array DDM Setup

© 2011 ANSYS, Inc. August 25, 201112

Examples1) Vivaldi Array

2) Probe Fed Patch Array

© 2011 ANSYS, Inc. August 25, 201113

Example #1: Vivaldi Antenna Array

256 element slant polarized Vivaldi array

Presenter

Presentation Notes

Using the same general element design as the 2010 finite array presentation, but added metal thickness for grounds and stripline to mimic real array. Also, added to element count. Last year was 144 elements, and this year is 256 elements with a more RAM intensive cell.

© 2011 ANSYS, Inc. August 25, 201114

Vivaldi Unit Cell Setup

PML

Master/slave boundaries on all outer walls

• Unit Cell consists of 4 Vivaldielements, 2 in each polarization

• Modeling an 8x8 array of the unit cell yields 256 elements

• Realistic 35um metal thickness for grounds; 18um for striplines

4 wave port excitations 2 for each polarization

© 2011 ANSYS, Inc. August 25, 201115

• Can solve large arrays with few cores. Here solved with 3 (minimum) and 13 cores.

• Solving > 11 million tets, using most accurate pattern settings, and with standard hardware!

• Fewer cores means less RAM, but more solve time. User can choose tradeoff.

Vivaldi Finite Array Using DDM Simulation Statistics

Model#

Sources#

CoresMeshTime

SolveTime

# TetsArray

Array TotalRAM

Max Domain

RAM

Avg. Domain

RAM

8x8 DDM Array 256 3 0h:7m 76h:18m 11,242,700 30.7GB 15.37GB 15.36GB


Presenter

Presentation Notes

Here we compare two machine list setups and the tradeoffs associated with the number of “machines” given in the list. 3 machines is the minimum number that HFSS will allow to solve with the array tool DDM. The algorithm always uses the first machine in the list as the head node, and all further machines as solve nodes. In the case of 3 “machines”, 2 of them are solve nodes, each node requiring an average of 15.36GB of RAM. With just 2 solve nodes, HFSS will solve 5 simultaneous domains per node, which is why the RAM total per node is higher with the 3 core list than with 13 core list. 15GB is ~3GB per unit cell times 5. In contrast, when the number of machines in the list is 9 or greater, the maximum number of simultaneous domains solved per node is 2, so a few of the machines were solving 2 simultaneous domains and some were solving 1. This is where the 4GB avg. RAM per domain comes from. Notice that the minimum number of simultaneous machines solves in the least amount of RAM and cores at the expense of a longer solve time. With more cores the solve time is over 2.5 times faster with a relatively smaller increase in RAM of 57%.

© 2011 ANSYS, Inc. August 25, 201116

Mutual Coupling

P1

Element A2

Element A1

Master/slave B direction

Master/slave A directionP2

P3 P41,1

2,1

1,2

1,1

1,3

3,11,2

2,3

3,3

Mutual coupling can be displayed in Designer’s Network Data Explorer in tabular or graphical format

Element B1 Element B2

P1 P2

P3 P4

P1 P2

P3 P4

P1 P2

P3 P4

P1 P2

P3 P4

P1 P2

P3 P4

P1 P2

P3 P4

P1 P2

P3 P4

P1 P2

P3 P4

Presenter

Presentation Notes

This slide shows the port numbering and port setup in each unit cell. The master/slave pairs determine the A and B directions along which the array is replicated. Exporting the .s256p file into touchstone format, we can import it into the network data explorer in Ansys Designer. Looking at the data table for the S-parameters is useful, but could be a bit daunting to quickly/easily see all of the data.

© 2011 ANSYS, Inc. August 25, 201117

Mutual Coupling with Network Data Explorer in Designer

0dB

-5dB

-10dB

-15dB

-20dB

-25dB

-30dB

-35dB

-40dB

• Simultaneous graphical RL and mutual coupling values• Set thresholds for coupling value range visualization• Quick identification of coupling ranges for large data sets

Presenter

Presentation Notes

Network data explorer allows graphically viewing the entire s-matrix with a user defined, color-coded legend. In this case, because there are multiple excitations per unit cell, the intuitiveness of the data explorer is reduced, but it can still be useful for quickly scanning Magnitudes that exceed a desired threshold. In this case the coupling threshold was set to -40dB and all values that are dark blue are below -40dB.

© 2011 ANSYS, Inc. August 25, 201118

Vivaldi Array Pattern Data

• Feeding only P1 element in each of four test cells.

• The center element has a more even pattern as might be expected

P1 P1

P1

© 2011 ANSYS, Inc. August 25, 201119

Explicitly Solved Finite Vivaldi Array vs. DDM Comparison

Vacuum buffer region mimics DDM

Model#

Sources#

CoresMeshTime

SolveTime

# TetsArray

Array TotalRAM

Max Domain

RAM

Avg. Domain

RAM


256 Element Explicit Array

256 12 Total = 122h:18m 5,881,409 211GB N/A N/A

Radiation boundary on sides

PML

Explicit 256 element slant polarized array

Presenter

Presentation Notes

The finite array was created in a similar manor as DDM would have created it. This shows the power of the Finite Array Domain Decomposition method. When the geometries get detailed and complicated the meshing process is simplified in such a way that the unsolvable becomes solvable.

© 2011 ANSYS, Inc. August 25, 201120

Inphi LRDIMM test machine– 284GB RAM

Special Thanks to Inphi

© 2011 ANSYS, Inc. August 25, 201121

Example #2: Probe Fed Patch Array

A. K. Skrivervik and J. R. Mosig, “Analysis of Finite Phased Arrays of Microstrip Patches,” IEEE Trans. Antennas Propagat., vol. 41, pp. 1105-1114, 1993.

Model a simple 9x5 probe-fed patch array from a reference

© 2011 ANSYS, Inc. August 25, 201122

Center Element Embedded Element Pattern Comparison with Reference

© 2011 ANSYS, Inc. August 25, 201123

Probe Fed Patch Array with Feed Network

Add complexity to the design to make it more realistic:

-7x7 = 49 elements

-Wilkinson power divider on stripline layer

-Via cage isolation on striplinelayers

Presenter

Presentation Notes

Array Description: Dual Probe, Single Polarization Patch Antenna with integrated Feed. Feed Consists of Wilkinson Power Divider and a 180deg line length added to one of the Wilkinson outputs to apply 180deg phase offset between probes. Patch is supported by a Rogers5880 substrate that is 1.57mm thick Feed Circuit is made of Stripline that has a ground plane spacing of 0.8433mm. The Feed’s substrate is Rogers R04232. The feed circuit is isolated with a via fence around the unit cell as well as throughout the trace routing. Bond Layer PrePregs of 3mils are also added between the board layers and all conductors are 0.7mil thick corresponding to 0.5oz copper. Shorting pin placed in the center of the patch to improve polarization isolation The lattice spacing is 14mm in the E-Plane and 13.77mm in the H-Plane. The array will be analyzed at 10GHz

© 2011 ANSYS, Inc. August 25, 201124

• Unit cell is analyzed to create mesh

• Unit cell mesh is then replicated to form remaining elements in the array

Finite Phased Array Setup Using DDM

Unit Cell Simulation Creates Mesh

Master / Slave Boundaries

Radiation Boundary

7x7 Finite Array

Presenter

Presentation Notes

Finite Phased Array DDM Setup

© 2011 ANSYS, Inc. August 25, 201125

• All mutual coupling terms are reported

• Here the coupling is reported from the center element to every other element in the array.

• Symmetry maintained down to the tenths of a dB level.

Mutual Coupling Data

-37dB

-34dB

-33dB

-31dB

-33dB

-34dB

-37dB

-33dB

-30dB

-29dB

-27dB

-29dB

-30dB

-33dB

-30dB

-24dB

-22dB

-26dB

-22dB

-24dB

-30dB

-33dB

-21dB

-13dB

-13dB

-13dB

-21dB

-33dB

-30dB

-24dB

-22dB

-26dB

-22dB

-24dB

-30dB

-33dB

-30dB

-29dB

-27dB

-29dB

-30dB

-33dB

-37dB

-34dB

-33dB

-31dB

-33dB

-34dB

-37dB

* Note: Mutual coupling terms does not show overlaid on top of geometry

Presenter

Presentation Notes

This slide shows the mutual coupling data between the enter element and all other elements in the array. It was created by looking at the solution data and manually writing on top of a picture of the array. Things to Note: The incredible symmetry across the array. It is exhibited vertically, horizontally and diagonally. This symmetry is accurate to the tenths of dB level. It is most likely due to the fact that the mesh itself is perfectly symmetric since we simply copy it from the unit cell simulation to each of the other domains. HFSS provides this data in the same manor as it does all other S-parameter data as it sees no difference between the two (it just solved the problem) Users can use this data to recreate the active return loss given different magnitude or phase tapers.

© 2011 ANSYS, Inc. August 25, 201126

• Active return loss is calculated by setting the desired excitations in the Edit Sources dialog

• Vertical symmetry maintained when array is scanned toward the left.

Active Return Loss with Phase Taper

-12dB

-9dB

-8dB

-10dB

-8dB

-9dB

-12dB

-12dB

-11dB

-9dB

-11dB

-9dB

-11dB

-12dB

-11dB

-11dB

-9dB

-10dB

-9dB

-11dB

-11dB

-10dB

-10dB

-8dB

-8dB

-8dB

-10dB

-10dB

-12dB

-10dB

-8dB

-9dB

-8dB

-10dB

-12dB

-15dB

-14dB

-12dB

-13dB

-12dB

-14dB

-15dB

-13dB

-18dB

-16dB

-16dB

-16dB

-18dB

-13dB

Presenter

Presentation Notes

This slide shows the active return loss when the array is scanned to a 45degree angle off boresite to the left of the slide. To accomplish this, a progressive phase taper was applied in that same direction. As a result, the vertical symmetry is maintained, but the horizontal symmetry is not. This occurs because we changed the phase of the signals excited on each of the elements which in tern changes the constructive and destructive interference ultimately affect the input impedance of the array. Similar affect could be seen with a magnitude taper as well, but probably would not show up as well because of the array’s size. This is essentially the return loss that would be seen when the array is excited in this manor. It was obtained by setting all the excitations with the appropriate phase and then plotting the ActiveS at each element. HFSS does not overlay this data on top of the geometry, but will report it in cartesian plots or data tables. To make filling out the Magnitude and Phase values easier a data file can be create and used to load in the values in one action.

© 2011 ANSYS, Inc. August 25, 201127

• The embedded element patterns predict expected behavior based on their location in the array.

Embedded Element Patterns

E-Plane Cut

Presenter

Presentation Notes

This radiation pattern shows E-Plane the embedded element pattern of 4 different elements in the array. The embedded element pattern (A.K.A Active Element Pattern) is the pattern of the array when a single element is excited and all other elements are match terminate. The E-plane is the plane that cuts vertically through the array. Note: 1. The center element’s E-plane pattern (Blue) is symmetric about theta = 0deg. This should be the case since it sees the same environment in either direction. The E-Plane edge element’s E-plane pattern (Green) is asymmetric about theta = 0deg. It is easiest to see this looking at the -60 and +60 degree angles in theta. We would expect this to be the case since this element sees other elements to one side in the E-plane and open space in the other side. The H-Plane edge element’s E-plane pattern (Red) is symmetrical about theta = 0deg. Although there are no elements to the right of this element we are not cutting the observation angle in this direction. In the E-plane this element sees the same environment regardless of whether it looks in the plus or minus theta direction. The corner element’s E-plane pattern (Black) is asymmetrical about theta =0deg. No matter what direction the radiation is observed this element sees a different environment which contributes to the asymmetry. Were we to look at the H-plane pattern we would also see an asymmetry as well.

© 2011 ANSYS, Inc. August 25, 201128

• Array has no phase shift taper so the beam is pointed toward boresite

• Sidelobe Levels are -10.86dB and -13.23 dB in the E & H Planes.

Array Pattern in the E-Plane

Presenter

Presentation Notes

This shows that the array can be scanned to boresite in a post-processing fashion by filling the proper equations into the Edit Sources Dialog window. The pattern is illustrated using a 3D polar plot superimposed on the array and through radiation polar plots of the E and H plane patterns. Notice: 1. The main beam points directly to 0 degrees in the E-plane pattern as specified. 2. The sidelobe level in both the E and H-plane pattern are -13.29dB which is fairly close to the expected -13dB expected for a large uniformly excited array.

© 2011 ANSYS, Inc. August 25, 201129

• Array has a progressive phase shift applied to the edit sources to scan the main beam to 45o in E-plane

• Actual beam steered to 43o due to the finite nature of the array

• Sidelobe Levels are -10.86dB and -13.23 dB in the E & H Planes.

Array Pattern Scanned to 45o in the E-Plane

Presenter

Presentation Notes

This shows that the array can be scanned off boresite in a post-processing fashion by filling the proper equations into the Edit Sources Dialog window. The steered beam is illustrated using a 3D polar plot superimposed on the array and through radiation polar plots of the E and H plane patterns. Notice: 1. The main beam points to 43 degrees in the E-plane pattern and not 45 degrees. This is typical of a finite array because of its size. The larger we make the array the less significant this beam squint becomes. 2. The sidelobe level in the H-plane pattern is -13.23dB which is fairly close to the expected -13dB expected for a large uniformly excited array.�3. The sidelobe level in the E-plane pattern is -10.86dB. This once again is due to the finite nature of the array and the affects of beam steering.

© 2011 ANSYS, Inc. August 25, 201130

Explicitly Solved Finite Patch Array vs. DDM Comparison

Vacuum buffer region mimics DDM

Model#

Sources#

CoresMeshTime

SolveTime

# TetsArray

Array TotalRAM

Max Domain

RAM

Avg. Domain

RAM


7x7 Array 49 8 Days ? ? ? N/A N/A

DDMExplicit finite array

Presenter

Presentation Notes

The finite array was created in a similar manner as DDM would have created it. Unfortunately after 3 days it still has not finished the initial mesh. This shows the power of the Finite Array Domain Decomposition method. When the geometries get detailed and complicated the meshing process is simplified in such a way that the unsolvable becomes solvable.

© 2011 ANSYS, Inc. August 25, 201131

• Instead of saving fields in the entire 3D volume, HFSS has a new option to save fields only on radiation surfaces.

• This drastically reduces disk space usage for problems such as large antenna arrays.

Efficiency Enhancements: Save Radiated Fields Only Option

ModelDisk

Space

Default: “Save Fields” (in 3D

Volume)15.6GB

With New “Save Radiated FieldsOnly” Option

0.66GB

9x5 patch array with 90 excitations

© 2011 ANSYS, Inc. August 25, 201132

Efficiency Enhancements to Edit Sources Dialogue

• The Edit Sources setup can now be saved to a file or loaded from a file in V14

• The 256 excitation Vivaldi array is shown here with only P1 excited for each element. This was done using a .csv file that was quickly edited in excel.

The excitation setup can be changed arbitrarily in post-processing.

© 2011 ANSYS, Inc. August 25, 201133

• Large finite antenna arrays are solved efficiently and quickly with the new array tool + DDM in HFSS V14

• The most accurate solution possible is achieved through mesh re-use of a well converged unit cell

• Setup of finite arrays is easier than ever, only requiring drawing of a single unit cell.

• The “Save Radiated Fields Only” option, and file based Edit Sources settings help to save substantial amounts of time and disk space

Conclusions

Documents

HFSS with HPC for Large Finite Antenna Array Design - Ansys · HFSS with HPC for Large Finite Antenna Array Design ... Solution: Finite Array Domain ... we can import it into the