A: Fitting GLE parameters from scratch is not for the faint hearted. Tersely documented code to do so is available from a public github repository . However, parameters for a given application can typically be transfered easily from a system to another. The input page provides a convenient interface to fetch pre-computed parameters from a library and adapt them to the application at hand.
A: Having small drift of the conserved quantity in a constant-temperature simulation is a sufficient but not necessary condition for accurately sampling the canonical ensemble. Optimal-sampling GLEs - just as aggressive white-noise thermostatting - interferes with the accuracy of integration and increases the drift of the conserved quantity. In general, however, this does not lead to measurable degradation of the accuracy of computed observables.
A: You just need to know an order-of-magnitude estimate of the maximum frequency in your system. Pick the value of ℏω/kT that approaches more closely this frequency: the colored noise will try to reproduce quantum effects for modes up to this cutoff. Choosing a value which is too high is not inherently problematic, but might require the use of a smaller time step, since the noise contains then very high frequency components.
A: Nothing is wrong. The "temperature" output by MD codes is typically just proportional to the instantaneous value of the classical kinetic energy, relying on the relation \[ \left\langle T\right\rangle=\frac{2}{3N_fk_B} \left\langle K\right\rangle. \] Clearly this relation only holds in a classical context. In a quantum-thermostat simulation the mean kinetic energy is (an approximation to) the quantum expectation value, that is not related in a simple way to the physical (ensemble) temperature. In a PI+GLE or PIGLET simulation, the particle momenta have no physical meaning, and there is no obvious meaning to the value of temperature printed by the code.