Ab initio, or first principles, methods solve the quantum mechanical equations which govern the behaviour of a system. The only information which must be provided are the atomic numbers and positions of the atoms within the system. In contrast, empirical or semi-empirical approaches require a model of the interactions between the atoms to be supplied. The parameters of these models are usually derived by fitting the outcome of simulations to experimental data. The problems with these techniques arise when you consider the question of the range of their applicability. If the parameters of the models were derived from system A, what guarantee is there that they apply to system B?
A number of approximations must be made in order to perform ab
initio calculations on all but the smallest of systems. However, as
will be seen in Chapter 
, these approximations are based
on general physical principles and do not rely on the specific nature
of the system under investigation. Hence, there is never a problem
about the assumptions of the model being violated by the changes made
during an investigation. This means that we may have more confidence
in the results of the `computational experiments' referred to in the
previous section. The use of ab initio techniques for predictive
calculations offers greater reliability for similar reasons.
The high computational cost of ab initio calculations implies that they are most profitably used for simulating the detailed electronic structure of a system or mechanism of a reaction. Due to the limitations in the size of system that may be addressed and the relatively high cost of ab initio calculations, there is still a rôle for semi-empirical methods in addressing questions on a larger scale. For example, a semi-empirical `Molecular Field Analysis' of interactions between a ligand and receptor may be applied to the `lock and key' model described in the previous section [1]. This allows rapid evaluation of molecules as potential ligands for a given receptor. As available computing power increases, the range of problems which may be addressed by ab initio methods will widen. At present, systems containing hundreds of atoms may be modelled on parallel supercomputers. The next generation of machines, with performance measured in TFLOPS (trillions of floating point operations per second), will allow the simulation of thousands of atoms.