Issues with 3rd order force constants in Silicon thermal conductivity calculation

hiPhive
1

I have been able to calculate phonon dispersions using hiPhive fitting on Quantum Espresso DFT. However, once I feed the second and third order force constants into ShengBTE, I get a thermal conductivity of ~96-100 instead of the expected 150 depending on what parameters I’ve used.

I did the convergence for plane wave cutoff and kpoint grid size and got 60 and 6x6x6, also converging the lattice constant at these parameters. I kept rattle_std at .01 and min_dist between 2.32 and 2.33. The memory in my PC limits the supercell dimensions to 3x3x3 for a primitive cell with 2 atoms and as such my cutoffs are limited to 5.87 for the second-order and set to 4.5 for third-order.

My fits are good (~.02 RMSE) and I’d gotten the forces to reasonable magnitudes after fiddling with rattle_std and min_dist (.6-.8 average with ~2 max). There are no errors or warnings from hiPhive or ShengBTE, but I am still about 1/3 off. I am changing to a higher convergence at 1e-8 from 1e-6 to increase accuracy in my next calculation (results TBD) but I wouldn’t think that would reduce scattering by that much. What are some other things that might be contributing to increased predicted scattering?

2

From what I remember in Si the thermal conductivity is quite sensetive to your cutoffs. And that specifically 2nd order cutoff needs to be quite large to recover the correct value.

So id suggest u do a careful convergence study of the conductivity with respect to cutoff and see if it helps.

Additionally having good amount of training structures with small isch displacements is likely a good starting point. Rattle structures of 0.6 sounds very large.

The thermal conductivity in Si w hiphive is carefully worked out in Efficient construction of linear models in materials modeling and applications to force constant expansions | npj Computational Materials

So you should be able to reproduce those results.