Self-adapting precision during optimizations

In Geometry Optimizations and Transition State searches, the gradients at the initial geometry may be quite large and in such case there is no need to apply the same high integration precision as may be required close to convergence when the gradients become very small. Therefore, adf tries to reduce the numerical integration accuracy during optimization to save time, and use the user-specified precision only in the final stage of the optimization.

The program starts with an initial value accfirst, to get an assessment of the (local) gradients. Subsequently it adjusts it according to the progress towards convergence. All values are kept between a lower and an upper bound: accmin and accmax respectively.

All three parameters, accfirst, accmin, accmax, are controlled by the key INTEGRATION.

INTEGRATION acc1 {acc2 {acc3}}

The simplest application, discussed above, specifies one value (acc1). This defines then both the upper bound (accmax) and the first value (accfirst). The lower bound accmin is by default 3.0 (adjusted internally to acc1, if that is lower); 3.5 in Transition State searches.

If two values are supplied, the smallest is taken for the lower bound, the larger for the upper bound. The value for the first cycle equals the upper bound.

If three values are supplied, the smallest is the lower bound, the largest the upper bound, the remaining value is used to start with.