The basic equation solved in a typical undamped modal analysis is the classical eigenvalue problem:
where:
| [K] = stiffness matrix |
| {Φi} = mode shape vector (eigenvector) of mode i |
| Ωi = natural circular frequency of mode i
(
|
| [M] = mass matrix |
Many numerical methods are available to solve the equation. Mechanical APDL offers these methods:
The damped and QR damped methods solve different equations. For more information, see Damped Method and QR Damped Method in the Mechanical APDL Theory Reference.
Table 3.5: Symmetric System Eigensolver Options
| Eigensolver | Application | Memory Required | Disk Required |
|---|---|---|---|
| Block Lanczos | To find many modes (about 40+) of large models. Recommended when the model consists of poorly shaped solid and shell elements. This solver performs well when the model consists of shells or a combination of shells and solids. | Medium | High |
| PCG Lanczos | To find few modes (up to about 100) of very large models (1,000,000+ degrees of freedom). This solver performs well when the lowest modes are sought for models that are dominated by well-shaped 3-D solid elements (that is, models that would typically be good candidates for the PCG iterative solver for a similar static or full transient analysis). | Medium | Low |
| Supernode | To find many modes (up to 10,000) efficiently. Use this method for 2-D plane or shell/beam structures (100 modes or more) and for 3-D solid structures (250 modes or more). | Medium | Low |
| Subspace | To find many modes (about 40+) of large models. This solver is more robust when the mass matrix is partially zero, or when there are u-P formulation elements in the model. It is recommended when using Distributed ANSYS. | Medium | High |
The Block Lanczos eigenvalue solver uses the Lanczos algorithm where the Lanczos recursion is performed with a block of vectors. The Block Lanczos method uses the sparse matrix solver, overriding any solver specified via the EQSLV command.
The Block Lanczos method is especially powerful when searching for
eigenfrequencies in a given part of the eigenvalue spectrum of a given system. The
convergence rate of the eigenfrequencies will be about the same when extracting
modes in the midrange and higher end of the spectrum as when extracting the lowest
modes. Therefore, when you use a shift frequency (FREQB
on MODOPT) to extract n modes
beyond the starting value of FREQB, the algorithm
extracts the n modes beyond
FREQB at about the same speed as it extracts the
lowest n modes.
While the Block Lanczos eigensolver is typically very robust for a wide range of applications, some element types may cause the eigensolver difficulty in achieving a final solution. These include MPC184 elements that use the Lagrange multiplier method and FLUID30, FLUID220, and FLUID221 elements that use the symmetric element matrix formulation for modal analyses (KEYOPT(2) = 2).
The PCG Lanczos method internally uses the Lanczos algorithm, combined with the PCG iterative solver. This method will be significantly faster than the Block Lanczos method for the following cases:
Large models that are dominated by 3-D solid elements and do not have ill-conditioned matrices due, for example, to poorly shaped elements
Only a few of the lowest modes are requested
Having ill-conditioned matrices or asking for many modes (e.g., more than 100 modes) can lead to an inefficient solution time with this method.
The PCG Lanczos method finds only the lowest eigenvalues. If a range of eigenfrequencies is requested on the MODOPT command, the PCG Lanczos method will find all of the eigenfrequencies below the lower value of the eigenfrequency range as well as the number of requested eigenfrequencies in the given eigenfrequency range. Thus the PCG Lanczos method is not recommended for problems when the lower value of the input eigenfrequency range is far from zero.
When to Use PCG Lanczos over Block Lanczos
The Block Lanczos eigensolver is the recommended eigensolver for most applications; however, the PCG Lanczos eigensolver can be more efficient than the Block Lanczos eigensolver for certain cases. Follow these guidelines when deciding whether to use the Block Lanczos or PCG Lanczos eigensolver.
Typically, the following conditions must be met for the PCG Lanczos eigensolver to be most efficient:
The model would be a good candidate for using the PCG solver in a similar static or full transient analysis.
The number of requested modes is less than a hundred.
The beginning frequency input on the MODOPT command is zero (or near zero).
The PCG Lanczos eigensolver (like all iterative solvers) is most efficient when the solution converges quickly. So if the model would not converge quickly in a similar static or full transient analysis, it is expected that the PCG Lanczos eigensolver will also not converge quickly and thus be less efficient. The PCG Lanczos eigensolver is most efficient when finding the lowest natural frequencies of the system. (For detailed information on measuring PCG Lanczos solver efficiency, see PCG Lanczos Solver Performance Output in the Performance Guide.) As the number of requested modes begins to approach one hundred, or when requesting only higher modes, the Block Lanczos eigensolver becomes a better choice.
Other factors such as the size of the problem and the hardware being used can
affect the eigensolver selection strategy. For example, when solving problems where
the Block Lanczos eigensolver runs in an out-of-core mode on a system with very slow
hard drive speed (that is, bad I/O performance) the PCG Lanczos eigensolver may be
the better choice as it does significantly less I/O, assuming a
Lev_Diff value of 1 through 4 is used
(PCGOPT command). Another example would be solving a model
with 15 million degrees of freedom. In this case, the Block Lanczos eigensolver
would need approximately 300 GB of hard drive space. If computer resources are
limited, the PCG Lanczos eigensolver may be the only viable choice for solving this
problem since the PCG Lanczos eigensolver does much less I/O than the Block Lanczos
eigensolver. The PCG Lanczos eigensolver has been successfully used on problems
exceeding 100 million degrees of freedom.
The Supernode (SNODE) solver is used to solve large, symmetric eigenvalue problems for many modes (up to 10,000 and beyond) in one solution. Typically, the reason for seeking many modes is to perform a subsequent mode-superposition or PSD analysis to solve for the response in a higher frequency range.
A supernode is a group of nodes from a group of elements. The supernodes for the
model are generated automatically by the program. This method first calculates
eigenmodes for each supernode in the range of 0.0 to FREQE*RangeFact (where RangeFact is
specified by the SNOPTION command and defaults to 2.0), and then
uses the supernode eigenmodes to calculate the global eigenmodes of the model in the
range of FREQB to FREQE (where FREQB and
FREQE are specified by the MODOPT
command).
Typically, this method offers faster solution times than Block Lanczos or PCG Lanczos if the number of modes requested is more than 200. The accuracy of the Supernode solution can be controlled by the SNOPTION command. For more information, see Supernode Method in the Mechanical APDL Theory Reference.
The lumped mass matrix option (LUMPM,ON) is not allowed when using the Supernode mode-extraction method. The consistent mass matrix option will be used regardless of the LUMPM setting.
When to Use Supernode
Typically, the following conditions must be met for the Supernode eigensolver to be most efficient:
The model is a good candidate for using the sparse solver in a similar static or full transient analysis (that is, the model is dominated with beams/shells or has a thin structure).
The number of requested modes is greater than 200.
The beginning frequency input on the MODOPT command is zero (or near zero).
For models that are dominated by solid elements or have a bulkier geometry, the Supernode eigensolver can still be more efficient than other eigensolvers; however, it may require higher numbers of modes to become the best choice in terms of performance. Other factors, such as the hardware being used, can also affect the decision of which eigensolver to use. For example, on machines with slow I/O performance the Supernode eigensolver may be the faster eigensolver (compared to Block Lanczos).
The Subspace method is an iterative algorithm that is
appropriate for problems having symmetric stiffness and mass matrices
( and
). It internally uses the sparse solver for the shift-invert logic.
The Subspace solver performs well when the goal is to obtain a moderate to medium
number of eigenvalues on large models run in distributed parallel mode.
Sturm sequence checking is available for this method (it is off by default) and can be controlled via the SUBOPT command. The memory usage of the algorithm can also be configured via the SUBOPT command.
The unsymmetric method, which also uses the
full and
matrices, is meant for problems where the stiffness and mass
matrices are unsymmetric (for example, acoustic fluid-structure interaction
problems). Sturm sequence checking is not available for this method. Therefore,
missed modes are a possibility at the higher end of the frequencies
extracted.
Solutions from the unsymmetric method can be real or complex. By default, the
solution type is determined automatically (Cpxmod = AUTO
on the MODOPT command).
For details about the complex frequency definition, see Complex Eigensolutions in the Mechanical APDL Theory Reference. For details about complex solution post-processing, see POST1 and POST26 – Complex Results Postprocessing in the Mechanical APDL Theory Reference.
The damped method
(MODOPT,DAMP) is meant for problems where damping cannot be
ignored, such as rotor dynamics applications. It uses the full matrices
(,
, and the damping matrix
).
Sturm sequence checking is not available for this method. Therefore, modes may be missing in the higher end of the range of extracted frequencies.
In a damped system, the eigensolutions are complex and the response at different nodes can be out of phase. At any given node, the amplitude will be the vector sum of the real and imaginary components of the eigenvector.
The complex frequencies are generally pairs of complex conjugates except in cases where system matrices are unsymmetric and structural damping is present.
For details about the complex frequency definition, see Complex Eigensolutions in the Mechanical APDL Theory Reference. For details about complex solution post-processing, see POST1 and POST26 – Complex Results Postprocessing in the Mechanical APDL Theory Reference.
The QR damped method (MODOPT,QRDAMP) combines the advantages of a symmetric eigensolution method with the complex Hessenberg method. The key concept is to approximately represent the complex damped eigenvalues by modal transformation using a larger number of eigenvectors of the undamped system. After the undamped mode shapes are evaluated by using the real eigensolution, the equations of motion are transformed to these modal coordinates.
Using the QR algorithm, a smaller eigenvalue problem is then solved in the modal subspace. This approach gives good results for lightly damped systems and can also apply to any arbitrary damping type or mix (proportional or non-proportional symmetric damping, nonsymmetrical gyroscopic damping matrix, and structural damping). This approach also supports nonsymmetrical stiffness if present in the model.
The QRDAMP eigensolver applies to models having an unsymmetrical global stiffness matrix where only a few elements contribute unsymmetric element stiffness matrices. For example, in a brake-friction problem, the local part of a model with friction contacts generates a unsymmetric stiffness matrix in contact elements. When an unsymmetric stiffness matrix is encountered, the eigenfrequencies and mode shapes obtained by the QRDAMP eigensolver should be verified by rerunning the analysis with the unsymmetric eigensolver. If an unsymmetric stiffness matrix is encountered, a warning message cautioning the user is output by the QRDAMP eigensolver right after the completion of the symmetric eigensolution.
The QRDAMP eigensolver works best when there is a larger “modal subspace” to converge and is therefore the best option for larger models. Because the accuracy of this method is dependent on the number of modes used in the calculations, a sufficient number of fundamental modes should be present (especially for highly damped systems) to provide good results. The QR damped method is not recommended for critically damped or overdamped systems. This method outputs both the real and imaginary eigenvalues (frequencies). By default, only the real eigenvectors (mode shapes) are output, although complex mode shapes can be calculated if requested with the MODOPT command.
The QRDAMP eigensolver is not recommended when system matrices are unsymmetric and structural damping is present. When these conditions are true, complex eigensolutions are not expected to be complex conjugate pairs.
Additional QRDAMP eigensolver options can be set via the QRDOPT command.
In general, ANSYS, Inc. recommends using the Damp eigensolver for small models. It produces more accurate eigensolutions than the QR Damped eigensolver for damped systems.
For details about the complex frequency definition, see Complex Eigensolutions in the Mechanical APDL Theory Reference. For details about complex solution post-processing, see POST1 and POST26 – Complex Results Postprocessing in the Mechanical APDL Theory Reference.