Any combination of methods for geometry optimization and wavefunction optimization is allowed. Possible options for geometry optimization are GDIIS, LBFGS, PRFO, RFO, BFGS and steepest descent. If you choose steepest descent for both, geometry variables and the wavefunction, a combined method is used. For all other combinations a full wavefunction optimization is performed between changes of the ionic coordinates. The convergence criteria for the wavefunction optimization can be adapted to the requirements of the geometry optimization (CONVERGENCE ADAPT and CONVERGENCE ENERGY). The default options are GDIIS and ODIIS. Some quasi-Newton methods (GDIIS, RFO and BFGS) are using the BFGS method to update an approximate Hessian. At the beginning of a run the Hessian can either be initialized as a unit matrix HESSIAN UNIT or with an empirical force field. Two force fields are implemented: The DISCO and the SCHLEGEL force field. The algorithm for the empirical force fields has to identify bonds in the system. For unusual geometries this may fail and the Hessian becomes singular. To prevent this you can add or delete bonds with the keyword CHANGE BONDS.
The linear-scaling geometry optimizers (options LBFGS and PRFO) do not require an approximate Hessian. To achieve linear scaling with the system size, the L-BFGS optimizer starts from a unit Hessian and applies the BFGS update on the fly using the history of the optimization. The P-RFO method can find transition states by following eigenmodes of the Hessian. The mode to be followed does not necessarily have to be the lowest eigenvalue initially (PRFO MODE). For larger systems, only the reaction core should be handled by the P-RFO optimizer (PRFO NVAR and PRFO CORE), and the environment is variationally decoupled using the L-BFGS optimizer. The normal LBFGS options can be used for the environment. The Hessian used for transition-state search therefore spans only a subset of all degrees of freedom, is separate from the Hessian for the other optimizers and the vibrational analysis, but it can be transferred into the appropriate degrees of freedom of the regular Hessian (HESSIAN PARTIAL). In order to allow negative eigenvalues, the Powell update instead of BFGS is used for transition-state search.
Although tailored for large systems, the linear-scaling geometry optimizers are suitable for smaller systems as well.