P.+nitx J. Ro.tni JoinN.i oi Fiiins ENoiNiiniNo Sir+ixnin 1994 Presented By: Steven Brus AME40510: Introduction to Numerical Methods May 4, 2011 S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 1 / 9 Qi.N+iiit.+ioN oi UNtin+.iN+v Calculations must be veried to determine their accuracy Vericationsolving governing equations right Accomplished through systematic grid renement studies Performance reports are often inconsistent and confusing Paper purposes a Grid Convergence Index (GCI) to uniformly report results of grid renements studies S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 2 / 9 Ritn.nnsoN Ex+n.roi.+ioN GCI method relies on the theory of Richardson Extrapolation Used to give an error estimate for a solution computed with a certain grid Requires two solutions: f 1 ne grid solution with grid spacing h 1 f 2 coarse grid solution with grid spacing h 2 Assumed series representation for discrete solution f f = f exact + g 1 h + g 2 h 2 + g 3 h 3 + . . . (1) Generalization for p th order methods and r valued grid ratio (h 2 /h 1 ) f exact f 1 + f 1 f 2 r p 1 (correction to ne grid) (2) f exact f 2 + (f 1 f 2 )r p r p 1 (correction to coarse grid) (3) S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 3 / 9 GCI ion Co.nsi .Nn FiNi Gnin Soii+ioNs Coarse Grid E coarse grid = ||r p r p 1 (4) GCI coarse grid = 3||r p r p 1 (5) Fine Grid E ne grid = || r p 1 (6) GCI ne grid = 3|| r p 1 (7) fractional deviation : = f 2 f 1 f 1 (8) S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 4 / 9 Asvxr+o+it R.Noi For the GCI to be valid, both grids must be in the asymptotic range To ensure this, 3 grids must be computed: f 3 (coarse) f 2 (intermediate) f 1 (ne) This yields 2 GCIs: f 1 , f 2 GCI 12 f 2 , f 3 GCI 23 The asymptotic range has been achieved if GCI 23 r p GCI 12 (9) log(e) log(h) 1 p S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 5 / 9 Ex.xrii Steady State Burgers Equation (for Re = 1000) u du dx + 1 Re d 2 u dx 2 = 0 u(0) = 1 u(1) = 0 (10) Second order centered dierences on a uniform grid, p = 2 Evaluate f = du dx at x = 1, r = 1.25 Fine grid (2000 cells): f 1 = 529.41 Coarse grid (1600 cells): f 2 = 544.48 Exact solution: f exact = 500 || = 100% f 2 f 1 f 1 = 2.85% (11) (r p 1) = (1.25 2 1) = 0.5625 (12) S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 6 / 9 Ex.xrii: FiNi Gnin Richardson Extrapolation error estimator E ne grid = || r p 1 = 2.85% 0.5625 = 5.07% (13) Grid Convergence Index GCI ne grid = 3|| r p 1 = 3 2.85% 0.5625 = 15.20% (14) Exact ne grid error A = 100% |f 1 f exact | f exact = 100% | 529.41 + 500.0| 500.0 = 5.88% (15) S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 7 / 9 Ex.xrii: Co.nsi Gnin Richardson Extrapolation error estimator E coarse grid = ||r p r p 1 = 2.85% 1.25 2 0.5625 = 7.92% (16) Grid Convergence Index GCI ne grid = 3||r p r p 1 = 3 2.85% 1.25 2 0.5625 = 23.75% (17) Exact coarse grid error A = 100% |f 2 f exact | f exact = 100% | 544.48 + 500.0| 500.0 = 8.90% (18) S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 8 / 9 CoNtiisioN GCI is a conservative error band whereas Richardson Extrapolation error estimator is not Allows the results of grid renement studies to be uniformly interpreted Non-integer grid renement can be used (r = 1.1 minimum) Can be applied a posteriori by editors and reviewers S+iviN Bnis (AME40510) On.i Riron+ M.v 4, 2011 9 / 9