Part 21 Report - 1996-042

 
 

ACCESSION #: 9606210210



ANSYS                    ANSYS, Inc.

                         201 Johnson Road         Telephone 412.746.3304

                         Houston, PA 15342-1300   Facsimile 412.746.9494



                         June 19, 1996



Thomas Greene, NRR/PECB

U.S. Nuclear Regulatory Commission

Mail Stop: 0-11E4

Washington, DC 20555-0001



Ref:      Accession 9601190322



Subj:     Request for a Technical Review of a Draft Information Notice

          Regarding Erroneous Results from ANSYS 5.0A and ANSYS 5.1

          Computer Codes



Dear Mr. Greene:



     ANSYS, Inc. takes issue with the Draft Information Notice; we

believe that it contains correct and misleading technical information.

We do not believe this condition is reportable as described since the

ANSYS numerical algorithms used in its solution are correct, The model

which was developed by Holtec is an idealization which cannot be

constructed in the physical world, and the results achieved by ANSYS are

completely consistent with the engineering idealization and numerical

methods employed.  With reference to the two points described to the

draft notice:



     1.   An explanation of this behavior was identified after Holtec's

          filing with the NRC, and is the agreed-upon explanation for the

          system-dependent solutions. Computer inaccuracy due to the

          idealized model and not ANSYS algorithms were the cause of the

          system-dependent solutions (see Holtec's fax dated June 19,

          1996).



     2.   To date, we have not received the model which was used to

          generate the results you describe as item 2 (i.e.  disparate

          results for forces and displacements are obtained depending

          upon the techniques employed to launch tho analysis runs).

          Your draft information notice was the first time we became

          aware of this claimed error.  Therefore, we cannot review and

          comment on the results until an input file is made available to

          us.  We have requested this information from Holtec and Mr.

          Y.S. Kim, NRR.



          In discussions with Holtec personnel, this behavior was never

          substantiated and may have been due to an aspect of their in-

          house submission procedure.  Until we review the ANSYS inputs

          and techniques employed, we cannot rule out user error.



                             illegible print



Thomas Greene

Nuclear Regulatory Commission

Page Two

June 19, 1996



Background



     A Holtec model was supplied to ANSYS, Inc. on August 7, 1995.  The

model had little damping in the gaps, severe nonlinearities, a path

dependent solution, and a large number of small time steps, Analyses were

performed using the Holtec model on two different Unix workstations which

resulted in similar solutions.  However, using Personal Computers the

solutions were different.  Recommendations were made in our October 18,

1995 letter to Holtec to modify their model so that these differences

would be minimized, Those modifications resulted in a less than 5%

difference in results from various Personal Computers.



     Further investigation of Holtec's model resulted in our realization

that the reported condition was due to Holtec's model relying on perfect

symmetry, The model was inherently unstable and prone to differences due

to machine precision, The model's perfect symmetry resulted in symmetric

and simultaneous contact with its enclosure, but only if the

displacements were perfectly symmetric.  A slight change in the models

dimensions resolves the problem immediately and results in repeatable

solutions on all platforms.



     Holtec now agrees that this is the phenomena which produced the

differing results and that the structure they modeled could never

physically exist.



     The way Holtec chose to model their problem made it inherently

numerically unstable and prone to differences due to machine precision.

The various hardware platforms on which we offer the ANSYS program have

differences in the fundamental precision at which the computations are

performed.  Various supported hardware and operating system environments

can provide as many as 80 bits of precision to 64 bits of precision.  Any

numerical representation with this many bits is more than sufficient for

engineering accuracy in a simple linear static analysis without numerical

complications.  However, when hundreds of thousands, or millions of

operations are repetitively performed on such numbers, the differences in

final results can begin to creep up into significant digits that may be

of interest to the engineer.



     For these reasons it is recommended that numerical sensitivity

studies be performed to determine the solution sensitivity to small

changes in the input.  If, for instance, small changes in geometric

locations (tolerances), material properties, convergence criteria or

damping result in large differences in the solution output, then it is

likely that tho modal or idealization is suspect.  On the other hand, if

small numerical changes in such input quantities have little effect on

the end result, then the sensitivity of the solution to numerical

precision can generally be eliminated as a concern.  The Holtec problem

is extremely sensitive.



Thomas Greene

Nuclear Regulatory Commission

Page Three

June 19, 1996



Detailed Description of the Holtec Model



     The structure being analyzed is a very stiff cross-braced space

frame, (approximately 70 in x 56 in x. 200 in high) represented in the

transient analysis by a superelement.  A superelement is a stiffness

matrix representation of a linear portion of the model.



     In the transient, the superelement (which has no constraints of its

own) is supported entirely by contact elements.  At the top, 4 sets of

contacts support +/- lateral motion with a gap of exactly 1.595 inches

each.  In reality, the structure is arbitrarily situated inside its

enclosure, but the easiest way to input the 1.595 clearance is to assume

a perfectly symmetric installation,



     The transient excitation is applied as a time-varying acceleration

by using the ACEL command in the x, y and z directions.  Since the

structure is symmetric and the gaps are identical, the ux and uy

displacements at the 4 top corner nodes are identical (for the displayed

number of digits) until some small numerical difference causes one of the

gaps to close at a fraction of a second before the others, The time at

which this occurs is a function of the machine accuracy.  It is only the

high accuracy of double precision arithmetic which allows the 4 sets of

contacts to close simultaneously for the first few collisions.



     This phenomenon has been demonstrated in a simpler model as a 2

dimensional block bouncing on a flat structure.  This model has been used

by ANSYS to demonstrate differences caused by computer precision.



     We request that this information notice he put on hold until these

technical issues have been resolved.  We are confident that the behavior

observed and reported is not an error in the ANSYS program.



                                        Sincerely,



                                        ANSYS, Inc.



                                        William J. Bryan

                                        Quality Assurance Manager



cc:  Holtec

          A. Soler

     ANSYS, Inc,

          D. Looman

          M. Imgrund

          J. Bittner

          D. Conover



*** END OF DOCUMENT ***