3 ms·

Relativistic DFT ... Brings back memories. I did non-relativistic DFT to model III-V materials in the early 90s. Basically building GaAs, InAs, etc. on superc

by hpcjoe 9y ago

Relativistic DFT ... Brings back memories. I did non-relativistic DFT to model III-V materials in the early 90s. Basically building GaAs, InAs, etc. on supercomputers using a set of codes we got from another research group, that we modified to do the calcs we needed. What I remember from this time was that our calculations were ridiculously intense because we could never get a machine with enough ram to store, or computational power to compute on the fly, all the matrix elements <phi_1|Operator|phi_2> we needed, so we wound up setting up huge tables on disk, and doing interpolation. My 64 atom supercells would take, initially a week of runtime to get 100 timesteps. After a few years, my SGI R8000 could do that calc about 90 seconds per time step. 10 years after I finished up, my home computer could do 10s wall clock per time step. I've not tried it recently, could be fun. Dirac eqn is a harder thing to deal with in general. We used Schrodinger eqn parameterized for soft pseudopotentials. Also had a restriction on how many k-points we could handle, and ionization was sadly impossible in the code, as the repeated volume of supercells made any unbalanced charge effectively infinite. I came up with some ways to handle this, but we lacked computer power for that. Likely it would work today. Honestly, I miss that work. It was fun.

plus 9y ago
> ionization was sadly impossible in the code, as the repeated volume of supercells made any unbalanced charge effectively infinite. I came up with some ways to handle this, but we lacked computer power for that. Calculations of charged periodic systems are possible through the use of Ewald summation techniques (or particle mesh Ewald/related techniques). It basically introduces a homogeneous neutralizing background charge, which is unphysical and makes it so that a the total energy of a charged particle in a box converges to the "true" result as a function of 1/volume. This is applicable to both classical and quantum simulations. There is no way to do a "true" charged periodic system, but honestly that's never what you want (because true infinite systems with net charge don't exist!). You can do better than homogeneous neutralizing background charge by introducing a neutralizing electrolyte through a continuum solvent model. Here's the implementation I'm most familiar with (https://arxiv.org/abs/1601.03346 https://arxiv.org/abs/1601.03346), but I know the fully open-source JDFTx also implements this (which has a bunch of other fancy approaches to incorporating solvation into periodic DFT calculations).

by hpcjoe 9y ago

plus 9y ago

> ionization was sadly impossible in the code, as the repeated volume of supercells made any unbalanced charge effectively infinite. I came up with some ways to handle this, but we lacked computer power for that. Calculations of charged periodic systems are possible through the use of Ewald summation techniques (or particle mesh Ewald/related techniques). It basically introduces a homogeneous neutralizing background charge, which is unphysical and makes it so that a the total energy of a charged particle in a box converges to the "true" result as a function of 1/volume. This is applicable to both classical and quantum simulations. There is no way to do a "true" charged periodic system, but honestly that's never what you want (because true infinite systems with net charge don't exist!). You can do better than homogeneous neutralizing background charge by introducing a neutralizing electrolyte through a continuum solvent model. Here's the implementation I'm most familiar with (https://arxiv.org/abs/1601.03346 https://arxiv.org/abs/1601.03346), but I know the fully open-source JDFTx also implements this (which has a bunch of other fancy approaches to incorporating solvation into periodic DFT calculations).