1 of 7

Improving DFT results

Abdul Muhaymin

Graduate student, MSN, Bilkent University

2 of 7

Outline

What to improve? What's wrong?

Even for 1-electron atom, this implies some kind of self-interaction which is unphysical
So: band gap overly underestimated (sometimes predicts things as metals while in fact they are semiconductor), LDA underestimate lattice parameters and hence overestimate bulk modulus
Answer: GGA?

LDA to GGA

Lattice parameters are really good in GGA
For structural parameters, PBEsol are especially good for solids
For adsorption energies (or material with hydrogen-bond effect), take a look at the revPBE (but it gives total energy a little bit higher)
So, GGA is overall better than LDA
However, little to no improvements in band gap

GGA to higher level of theory

meta-GGA functional gives very accurate band gap but computationally very inefficient
LDA, GGA, meta-GGA shows a systematic improvement, there is this accuracy-computational resource trade-off
Hybrid is even more computationally inefficient but a popular choice due to availability
Maybe check DFT-½ method which claims to be as efficient as PBE but as accurate as hybrid functional method (for semiconductors)
And remember explaining the physics is important!

DFT+U

All the time we are talking about DFT, all XC functionals are in some sense originated from HEG which is completely delocalized
So strongly-correlated systems or system where electrons are localized cannot be described well by standard DFT (such as d- and f- orbital transition metals)
That's why we use a correction term to DFT

The Hubbard parameter, U

U can be fine-tuned to match the experimental band gap (but the first-principle nature is now contaminated)
U can be calculated from linear response theory – hp.x, ACBN0, cRPA
There are other terms too: DFT+U+V, DFT+U+J

PDOS of CoO with regular DFT