Which one is the best, reliable and efficient DFT simulation package (with and without hybrid functionals) other than VASP which could calculate all possible properties of the materials?
I am interested in exploring the electronic properties of 2D materials (graphene, silicene, MoS2, etc), their van der waal hetero-structures and hybrid structures with DFT (TD-DFT and MD) calculations. Both free and paid license codes are acceptable but it is hard to pay for a expensive code. The computational resources would be up to 8 nodes = 96 cores. In addition to DFT simulation package, please add your suggestions with respect to plane-wave DFT vs local orbitals or standard DFT vs linear-scaling DFT with advantages and disadvantages. Thank you.
Now we can converge towards a more meaningful answer. I'm sure other participants will have experience with a wider variety of DFT codes to suggest good choices for the systems you want to study. My advice would be to do a quick literature review of works on similar systems and see what others in the field are using.
Meanwhile I'll try to briefly comment on your final questions.
Two main classes of standard DFT.
Plane-wave ("physicist's") DFT:
- expands the KS orbitals in terms of plane waves, up to a certain max kinetic energy ('KE cutoff'),
- heavily relies on performing certain stages in reciprocal space (FFTs), and thus traditionally assumes your system is fully periodic ,
- perfect for crystals, where extremely small unit cells can be used, and k-point sampling can be utilised,
- more tricky in the absence of periodicity in one or more directions. You can still pretend your system is periodic by padding it with vacuum (in your case -- only in one direction), but this is easier said than done. For caveats and overview of approaches to mitigate, see [1]. Mixed periodicity (like in your case) is slightly more tricky still, might want to look at [2]. Many, but not all, PW-DFT codes implement at least some of the approaches discussed in [1]. You might want to check that your code of choice does,
- the choice of plane waves as a basis set means that one usually "pays for vacuum" (more padding means more computational effort and higher memory requirements),
- plane waves do not move with atoms, meaning there are no Pulay forces (makes MD and geometry optimisation easier),
- similarly, no basis-set superposition error (BSSE),
- PW basis set depends on KE cutoff, and thus on grid spacing. This makes varying the simulation cell size or shape tricky, although not impossible. In many applications you don't need to vary the box size or shape. Notable exceptions: NpT MD, geometry optimisation of crystals.
- Pseudopotential approximation almost always necessary, because the all-electron wavefunction oscillates rapidly close to the nucleus. Use of pseudopotentials introduces subtle or not-so-subtle transferability issues. The nasty effects of pseudopotentials can be mitigated with PAW (worth checking if your code supports it).
- usually costs O(N^3) with the number of atoms.
- cost of using hybrid functionals comparatively high.
- (some) notable examples: CASTEP, Quantum espresso, VASP, CPMD, ABINIT.
Localised-orbital (typically Gaussian) ("chemist's") DFT:
- uses strictly-localised or (more often) fast-decaying orbitals,
- more freedom in the choice of boundary conditions, as periodicity is usually not assumed,
- usually no or comparatively little cost for vaccuum,
- basis functions move with atoms, leading to Pulay forces. This makes MD and geometry optimisation more difficult (but it's mostly a worry for the developers),
- basis-set superposition error needs to be taken into account (important!),
- free of the pseudopotential approximation,
- usually costs O(N^3) with the number of atoms.
- cost of using hybrid functionals comparatively low.
- likely to be faster than PW-DFT for small systems, especially if using hybrid functionals or if satisfied with lower-quality basis sets.
- cost very quickly increases once you need high accuracy.
- (some) notable examples: NWCHEM, G****ian, ADF, TURBOMOLE.
Not all approaches fit neatly into this classification, e.g. there exist strictly-periodic local orbital approaches, or mixed approaches.
Now for standard DFT vs linear-scaling DFT:
Standard DFT:
- almost universally scales as O(N^3),
- this means that with 96 cores you'll have a difficult time working with more than 300-500 atoms at reasonable quality (for singlepoint energy calculations).
- for geometry optimisation, half of that, probably,
- for TD-DFT and (particularly) DFTMD -- only very small systems will be practical, which may be problematic,
- however, standard DFT has been around for a long while, and the codes are more mature, so you can expect more functionality and properties (spectra, etc).
Linear-scaling DFT:
- scales as O(N) for large systems, between O(N^3) and O(N) for small and medium-size systems,
- for small systems (roughly 50 atoms), it will not be faster than standard DFT,
- for single-point energy calculations you should be able to do 2000-5000 atoms with 96 cores.
- similarly smaller system sizes admissible to geopt, TD-DFT, DFTMD.
- introduces additional approximations on top of DFT. While accuracy is usually well-controllable, meaning you can get arbitrarily close to the "true-DFT" result, this requires some experience. It's much easier to get garbage with LS-DFT than DFT if you are not careful.
- has seen only about 25 years of development to date, so codes are generally less mature.
- notable examples: ONETEP, SIESTA, CONQUEST, BigDFT.
[1] Hine et al, Electrostatic Interactions in Finite Systems treated with Periodic Boundary Conditions: Application to Linear-Scaling Density Functional Theory, J. Chem. Phys. 135, 204103 (2011).
[2] Leslie and Gillan, Journal of Physics C: Solid State Physics, 18 (5).
Dear Muhammad,
Admittedly your question is quite wide and admits many answers. First of all, since you discard VASP I assume you want something that is free and open source.
Secondly, you must tell us what kind of systems you want to study.
For periodic systems you should consider codes based on plane waves methods. Then ABINIT and Quantum Espresso are perhaps the tools you might look at. For molecules, you should look at real space codes, and one possibility is ELK (although it is indeed a mixed approach based on Augmented Plane Waves).
All these codes are quite well tested and in continuous expansion.
I am sure on the other hand that people from the forum will provide more answers offering a wider point of view.
Cheers,
Roberto
Insufficient data for meaningful answer.
You need to specify your requirements in more detail. There is no single best package, just like there is no single best tool in a toolbox. Is a hammer better than a screwdriver? Depends on what you're trying to accomplish.
Are your systems of interest strictly periodic? If so, in all dimensions? Do your requirements include isolated (aperiodic) molecules? Are you planning to do DFTMD or geometry optimisation? If so, do your requirements include varying the size of the simulation cell? Do your systems have a band gap, or can they be metallic? Is there strong correlation involved? What kinds of chemical elements are you planning to handle, anything exotic? What system sizes do you have in mind -- are they bigger than, say, 300 atoms? What computing resources do you have at your disposal -- 10 CPU cores?, 100?, 1000? Are you interested in free software only, or can you pay for a license? If so, what is your budget? Most importantly of all, what does "all possible properties" entail exactly? No code can calculate that.
Without even a rough idea of what you are trying to accomplish, it's difficult to even decide on the *classes* of approaches -- e.g. plane-wave DFT vs local orbitals or standard DFT vs linear-scaling DFT.
Thank you for adding your answers. I am interested in exploring the electronic properties of 2D materials (graphene, silicene, MoS2, etc), their van der waal hetero-structures and hybrid structures with DFT (TD-DFT and MD) calculations. Both free and paid license codes are acceptable but it is hard to pay for a expensive code. The computational resources would be up to 8 nodes = 96 cores. In addition to DFT simulation package, please add your suggestions with respect to plane-wave DFT vs local orbitals or standard DFT vs linear-scaling DFT with advantages and disadvantages. Thank you.
Now we can converge towards a more meaningful answer. I'm sure other participants will have experience with a wider variety of DFT codes to suggest good choices for the systems you want to study. My advice would be to do a quick literature review of works on similar systems and see what others in the field are using.
Meanwhile I'll try to briefly comment on your final questions.
Two main classes of standard DFT.
Plane-wave ("physicist's") DFT:
- expands the KS orbitals in terms of plane waves, up to a certain max kinetic energy ('KE cutoff'),
- heavily relies on performing certain stages in reciprocal space (FFTs), and thus traditionally assumes your system is fully periodic ,
- perfect for crystals, where extremely small unit cells can be used, and k-point sampling can be utilised,
- more tricky in the absence of periodicity in one or more directions. You can still pretend your system is periodic by padding it with vacuum (in your case -- only in one direction), but this is easier said than done. For caveats and overview of approaches to mitigate, see [1]. Mixed periodicity (like in your case) is slightly more tricky still, might want to look at [2]. Many, but not all, PW-DFT codes implement at least some of the approaches discussed in [1]. You might want to check that your code of choice does,
- the choice of plane waves as a basis set means that one usually "pays for vacuum" (more padding means more computational effort and higher memory requirements),
- plane waves do not move with atoms, meaning there are no Pulay forces (makes MD and geometry optimisation easier),
- similarly, no basis-set superposition error (BSSE),
- PW basis set depends on KE cutoff, and thus on grid spacing. This makes varying the simulation cell size or shape tricky, although not impossible. In many applications you don't need to vary the box size or shape. Notable exceptions: NpT MD, geometry optimisation of crystals.
- Pseudopotential approximation almost always necessary, because the all-electron wavefunction oscillates rapidly close to the nucleus. Use of pseudopotentials introduces subtle or not-so-subtle transferability issues. The nasty effects of pseudopotentials can be mitigated with PAW (worth checking if your code supports it).
- usually costs O(N^3) with the number of atoms.
- cost of using hybrid functionals comparatively high.
- (some) notable examples: CASTEP, Quantum espresso, VASP, CPMD, ABINIT.
Localised-orbital (typically Gaussian) ("chemist's") DFT:
- uses strictly-localised or (more often) fast-decaying orbitals,
- more freedom in the choice of boundary conditions, as periodicity is usually not assumed,
- usually no or comparatively little cost for vaccuum,
- basis functions move with atoms, leading to Pulay forces. This makes MD and geometry optimisation more difficult (but it's mostly a worry for the developers),
- basis-set superposition error needs to be taken into account (important!),
- free of the pseudopotential approximation,
- usually costs O(N^3) with the number of atoms.
- cost of using hybrid functionals comparatively low.
- likely to be faster than PW-DFT for small systems, especially if using hybrid functionals or if satisfied with lower-quality basis sets.
- cost very quickly increases once you need high accuracy.
- (some) notable examples: NWCHEM, G****ian, ADF, TURBOMOLE.
Not all approaches fit neatly into this classification, e.g. there exist strictly-periodic local orbital approaches, or mixed approaches.
Now for standard DFT vs linear-scaling DFT:
Standard DFT:
- almost universally scales as O(N^3),
- this means that with 96 cores you'll have a difficult time working with more than 300-500 atoms at reasonable quality (for singlepoint energy calculations).
- for geometry optimisation, half of that, probably,
- for TD-DFT and (particularly) DFTMD -- only very small systems will be practical, which may be problematic,
- however, standard DFT has been around for a long while, and the codes are more mature, so you can expect more functionality and properties (spectra, etc).
Linear-scaling DFT:
- scales as O(N) for large systems, between O(N^3) and O(N) for small and medium-size systems,
- for small systems (roughly 50 atoms), it will not be faster than standard DFT,
- for single-point energy calculations you should be able to do 2000-5000 atoms with 96 cores.
- similarly smaller system sizes admissible to geopt, TD-DFT, DFTMD.
- introduces additional approximations on top of DFT. While accuracy is usually well-controllable, meaning you can get arbitrarily close to the "true-DFT" result, this requires some experience. It's much easier to get garbage with LS-DFT than DFT if you are not careful.
- has seen only about 25 years of development to date, so codes are generally less mature.
- notable examples: ONETEP, SIESTA, CONQUEST, BigDFT.
[1] Hine et al, Electrostatic Interactions in Finite Systems treated with Periodic Boundary Conditions: Application to Linear-Scaling Density Functional Theory, J. Chem. Phys. 135, 204103 (2011).
[2] Leslie and Gillan, Journal of Physics C: Solid State Physics, 18 (5).
Thank you very much for your comprehensive answer in the selection of DFT codes.
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Materials Science
- Materials
- Carbon Materials
- Graphite
- Graphene