any specific reason that smartweights is not recommended right now?
~how to handle inconsistent user input? (fitting to stresses = 0 but still fits to stresses! true in FitSNAP too?)~
LH sampling is a bit spread out right now. HpStruct instance calls its method lh_params, which seems to set up a dictionary, then calls the libmod function latin_hypercube_sampling. i think this is to allow the simulated_annealing to also access LH sampling without forcing the FitSNAP sampling scheme into it. but might be good to confirm.
is purpose of initial fit in script_optimize just to get the descriptors calculated (i.e. process_configs)? if so, we should handle the initial fit within libmod_optimize.
related: how does readin in the initial weights do anyhting?