What steps can I take to improve model performance in order to ameliorate problems with bad derivatives?
In some cases there are steps that you can take to considerably increase the chances that model outputs will be differentiable with respect to parameter values. Here are a few basic (but very important) suggestions. Remember when considering these, that finite-difference derivatives (as calculated by PEST) are computed by subtracting one (possibly large) number from another (possibly large) number. This is an operation that is easily contaminated by numerical error. • If the model has an iterative solver, set the convergence criterion for that solver to a small value. • If the model uses adaptive time stepping, consider setting the model time step manually to an interval that is smaller than any which the model’s adaptive time-stepping scheme is likely to undercut. • Make sure that the model writes to its output files with maximum precision. This is between 6 and 7 significant figures for single precision arithmetic. Use exponential notation rather than fixed field notation for writing
Related Questions
- What steps can I take to improve model performance in order to ameliorate problems with bad derivatives?
- What options are available in PEST for improving its performance in the face of bad model derivatives?
- What options are available to me if a model’s derivatives are so bad that PEST simply will not work?
