3.5.3 Automatic Local Scaling



next up previous contents
Next: 3.5.4 Solving the Linear Up: 3.5 Solving the Nonlinear Previous: 3.5.2 Fréchet Derivative (Jacobian

3.5.3 Automatic Local Scaling

 

We solve the nonlinear system by constructing a sequence of improved solutions with the residuals . Improved means that the residuals' norms decrease with each update (). The nonlinear system is assumed to be solved when the residual norm is sufficiently small (3.5-9).

 

The termination criterion (Newton accuracy) has to be consistent with the magnitude of the terms in . For simulators where the equations are known a priori, e.g. for device simulators, it is possible to find an appropriate scaling from physical considerations. In PROMIS we do not make any restrictions to specific equations or physical problems and, therefore, scaling is based purely on numerical considerations.

We have incorporated a local scaling for the system . For each equation at each grid point we apply a local scaling factor (3.5-10) derived from (3.5-7).

 

 

The variables themselves remain unscaled and, therefore, in addition to the right hand side , the rows of the Jacobian matrix are scaled with the identical factors (3.5-12).

 

A consequence of the local scaling is that all equations are solved to the same degree of accuracy at all grid points independent of the magnitude of the involved physical quantities. In the further text, we omit the explicit indication of the scaled equations and use the same symbols irrespective on whether they are scaled or not.



Martin Stiftinger
Wed Oct 19 13:03:34 MET 1994