Why do density-functional corrections capture electron correlation better than fixed empirical potentials at bond-breaking geometries?

Your question asks about a fundamental gap in electronic structure theory: why do density-functional methods outperform simple empirical potentials when molecules break apart, where electron correlation becomes especially tricky.

The core issue is that fixed empirical potentials treat electron-electron interactions in a static, averaged way. When a bond breaks, the electron density redistributes dramatically — what was a shared pair of electrons now becomes two separate ones on distant atoms. An empirical potential calibrated to equilibrium geometry can't adapt to this new regime. Density-functional approaches, by contrast, adjust the effective potential *self-consistently* based on the evolving electron density [AG-2024.05-2255].

The H₂ molecule is the canonical test case. At equilibrium, both methods might work equally well. But as the bond stretches, standard DFT functionals (like the SCAN functional) remain accurate across all geometries precisely because they encode a response to changes in electron density [AG-2024.05-2255]. A remarkable recent result shows that even a *simple analytical DFT expression* — derived from the "independent atom ansatz," which treats the two atoms as nearly separate entities — recovers 99.5% of the correlation energy and correctly dissociates the bond with only 0.19 eV error [AG-2024.05-2255]. This works because the formula explicitly captures the Heitler-London resonance (the quantum tunneling tendency of electrons to hop between atoms) without needing empirical fitting [AG-2024.05-2255].

Why does this matter? Because bond-breaking is the gateway to chemistry. Any method that fails at dissociation cannot reliably predict reaction barriers, photochemistry, or excited-state dynamics. The self-consistent density response in DFT gives it adaptive power that fixed potentials lack.

That said, the preprints provided focus mainly on analytical expressions and multi-configurational methods rather than a systematic comparison of DFT against empirical potentials at bond-breaking geometries — so they don't directly address your specific question.