Two reference energies are available, HartreeFock, which corresponds to electrons moving independently of each other in a meanfield approximation [50]; and an exact energy obtained from a parameterization [26] of a quantum Monte Carlo calculation [25]. By taking the difference between these two energies, we compute a correlation energy, which the original eFF made no effort to incorporate, but which is necessary to include to achieve our ultimate goal of chemical accuracy.
We define an exchange energy as a pairwise sum over samespin electrons, and a correlation energy as a pairwise sum over oppositespin electrons. For the uniform electron gas, we use exchange and correlation functions defined as:
Using these simple pairwise functions, we are able to reproduce the equation of state curves of both the HartreeFock and exact uniform electron gas systems (Figure 35).

We find the energies of different crystalline packings to be similar ( hartrees/electron), suggesting that the uniform electron gas under eFF is highly fluxional (Figure 36).
To address this discrepancy, we add on an extra term that is dependent on the one particle exchange energy , and adjust its weighting to reproduce the dimer potential energy surface.
We also add on a term that depends on the ratio of electron sizes, which is zero in the uniform electron gas, but may make a contribution in molecules containing dissimilar nuclei. We adjust its weighting to reproduce the correct bond length of lithium hydride.
The final exchange functional for systems with nuclei is as follows:
With this functional we are able to reproduce with high accuracy the HartreeFock geometries and energies of simple compounds containing atoms from Z=16 (Figure 38, 39). No effort has been made yet to combine eFF2 and eFF3 into a single force field that could handle s and p electrons with high accuracy, but it would be a natural next step.

However, more work is needed, as seen when we try to reproduce correlation energies for the systems studied in the previous section (Figure 41).
We note that this method of computing correlation in principle carries some advantages over density functional theory, as it is an inherently nonlocal description [53]. It may be possible, for example, to achieve a good description of van der Waals forces by adding additional pairwise terms to the eFF correlation functional.