Method to Increase Computing Speed and Maintain Accuracy
AMS (Advanced Mass Scaling) saves significant computation time by increasing the time step of the model for an explicit computation. This is similar to traditional mass scaling, except that the added mass does not increase the translational kinetic energy of the system.
A non-diagonal mass matrix is used to increase the time step on each line of the mass matrix. The lumped mass, M0, is increased with some M value compensated with non-diagonal terms such that the total mass to remain constant. Unlike traditional mass scaling, AMS only modifies high frequencies and does not significantly affect low frequencies of the model.
The advantage of AMS versus traditional user controlled mass scaling is that translational kinetic energy is not increased. This allows the time step to be increased to significantly higher values as compared to traditional mass scaling without significantly affecting the results quality.
Since AMS does not modify the global mass, even at large time steps, the global momentum of the nodes affected by AMS is conserved. At large time steps, traditional /DT/NODA/CST can add a significant amount mass to a computation which increases translational kinetic energy.
AMS has a computational cost associated with calculating the mass matrix. The computational cost is model dependent, but for a highly nonlinear model it could be 50% of the total computational cost. So, although the cost per cycle has increased, the number of calculation cycles is reduced, due to the increased time step. For example, using a time step of 10 times the traditional /DT/NODA/CST, the total time for the calculation was reduced by a factor of 3. Therefore, to see a reasonable reduction in elapsed time, 10 times the /DT/NODA/CST time step is the recommended starting point.
Computational convergence and accurate results can be obtained by setting a target time step to 10 to 20 times higher than traditional mass scaling. In manufacturing simulations, 50 times the traditional mass scaling time step can be used. Since the Courant condition remains to be respected, the stability of the model must be achieved with the targeted time step to apply AMS.
Several modifications in the model may help increase its stability with high time step. Below are some recommendations and suggestions in order to insure the stability of the model.
| Note: | Advanced Mass Scaling is specific to RADIOSS. It is advanced because it can be applied to the entire model without degrading computing performances and result quality. |
|
What is new about AMS in Version 14.0
| • | Compatibility with RBE2 and RBE3 |
What was new about AMS in Version 13.0
| • | Compatibility with moving rigid walls (/RWALL with node_ID > 0) |
| • | Fixed rigid walls were corrected (/RWALL with node_ID = 0 or blank) |
| • | Tolerance default value was changed from 1E-4 to 1E-3 (Tol_AMS = 0  0.001) |
| • | Conjugate Gradient (C.G.) convergence criteria was improved |
| • | Non-diagonal added mass matrix was optimized |
| Note: | The AMS tolerance was changed in order to compensate a slight loss of computation time performance, due to above listed improvements but it should not affect the results accuracy. |
|
Starter
| • | Only Keyword: only /AMS must be present in order to apply the AMS to either a designated group of parts, or the entire model, if followed by a blank line (RADIOSS Starter Input). |
| Caution: | /DT/AMS is mandatory in RADIOSS Engine Input followed by the scale factor and the targeted increased time step to activate AMS (RADIOSS Engine Input). Without /DT/AMS in the Engine file, /AMS in the Starter file is ignored. |
| • | HMPP / SPMD: if DOMDEC is set to 0, the Starter points to 3 (Multi-level Kway decomposition) by default.
Up until version 11.0.230, if /AMS is applied, DOMDEC must be manually switched to 5 (DOF-based Multi-level Kway decomposition).
As of version 11.0.240, the Starter automatically sets DOMDEC to 5, if it is set to 0 and if /AMS is present, otherwise, DOMDEC will be set to 3, if /AMS is absent. |
| - | It is recommended to set Interface stiffness flag, Istf to 4 (K=min(Km,Ks) and Stiffness scale factor, Stfac at its default value of 1 for contact interfaces TYPE7, TYPE11, TYPE19, and TYPE20. |
| - | If friction is involved, it is advised to set the friction penalty formulation type, Iform to 2 (incremental stiffness for Coulomb friction), whenever this option is available (/INTER/TYPE7, /TYPE19, and /TYPE20). |
| - | For AMS, like in standard mass scaling, it is recommended to not have friction in a TYPE11 contact, if a TYPE7 (already handling friction) contact is already defined for the same parts. This avoids drops of time step and helps model convergence. This recommendation becomes obsolete, if TYPE11 friction is using Iform=2 for edge-to-edge contacts, as of version 13.0. |
| - | With contact interfaces TYPE7, TYPE11, TYPE19, and TYPE20 using nonlinear penalty stiffness for contact, it might be necessary to use the /DT/INTER/DEL option in the RADIOSS Engine input deck. Otherwise, AMS may slowly converge, or may diverge. |
| - | Using /DT/INTER/AMS is meaningless when using /DT/AMS, but /DT/INTER/AMS can be used, instead of /DT/NODA/CST when contacts highly affect the time step. |
| - | Initial intersections and penetrations should be removed, as much as possible (as well as in traditional mass scaling, notably if it is used for comparison with the AMS results) |
| Caution: | Barriers, other impactors, and more particularly dummies contain a lot of interfaces that must NOT be modified since their validation is based on their original /INTER options (if modified, their validation is not guaranteed). |
| - | Small rigid bodies inertia should rather be spherical with Ispher set to 1 |
| Caution: | Dummies contain a lot of small rigid bodies that must NOT be modified since their validation is based on their original /RBODY options (if modified, their validation is not guaranteed). |
| • | Flying nodes: are not a limitation but they should be removed, as well as unconnected small rigid bodies that are not needed. |
| • | Rayleigh damping: in few cases where the elastic domain remains large along the simulation, poor AMS performances with noisy time history curves (example, force-time) and Harlequin patterns in contour plots (example, Von Mises) may be observed. A significant improvement should occur by inserting a Rayleigh damping card (/DAMP in the Starter file) affecting the parts subject to AMS using these parameters:  =0 and  =0.05*ΔtAMS ( is 5% times the AMS target time step). Such a modification may allow to increase the AMS target time step for much better AMS performances.
Note: it is not advised to use numerical damping (dm or Navier-Stokes), instead of Rayleigh damping, here. |
Engine
| • | Only Keyword: only /DT/AMS must be present in order to invoke AMS with the targeted increased elementary time step (RADIOSS Engine Input). |
| Caution: | /AMS is mandatory in the RADIOSS Starter Input in order to designate the selected parts or the entire model on which AMS is applied. |
| • | In /DT/AMS/Iflag if Iflag = 1, the tolerance for AMS convergence, Tol_AMS, must be provided. If Iflag is ignored, the default value (1E-3) for Tol_AMS is applied. It is not recommended to modify AMS tolerance, if it is not needed. Note that Tol_AMS was 1E-4 in 12.0.210. |
| • | In /DT/AMS/Iflag if Iflag = 2, the display of the number of AMS iterations may help in debugging or monitoring convergence at no extra cpu cost. Maximum allowed iterations before sending a divergence message is 1000 by default. It is advised to NOT change this value. Anything above this maximum value, RADIOSS stops with the following error message: |
** ERROR ** AMS IS LIKELY DIVERGING
Monitoring the number of iterations per cycle can help to understand the resulting AMS performance:
75 to 100 iterations is a sign of a really bad convergence
50 iterations may not provide any speedup
30 iterations or less is considered a good convergence.
| • | It is advised to use 0.67 as the scale factor, ΔTsca, rather than 0.9 (as well as in traditional nodal mass scaling /DT/NODA/CST, especially if it is used for comparison with the AMS results). |
| • | Target a minimum time step, Tmin, from 10 to 20 times than the one used in /DT/NODA/CST e.g. start with 10 times, then check how far to go by verifying the numerical convergence, the gain of elapsed time, and the results quality. |
| • | /DT/AMS will not modify the global mass, therefore: |
| - | The global momentum of the related nodes is conserved. |
| - | A comparison with traditional mass scaling requires a minor added mass along the /DT/NODA/CST computation, if used as a reference. |
| • | Use /DT/INTER/DEL as a workaround for slow convergence, or possible divergence, when invoking a nonlinear penalty stiffness with contact interfaces TYPE7, TYPE11, TYPE19, and TYPE20. |
| • | /ANIM/NODA/NDMAS may be added in order to output non-diagonal mass variation. |
| • | /ANIM/NODA/DINER may be added in order to output added inertia per nodes. |
See Also:
/AMS
/DT/AMS
