Read on Twitter
Happy to join #VCT2020 with my poster #VCT2020EU_P39. It shows how machine-learned interatomic potentials can be used to accelerate the prediction of phonon properties. Work with @vl_deringer.
Publication soon in #AIP_JCP of @AIP_Publishing
Preprint: https://arxiv.org/abs/2005.07046
Publication soon in #AIP_JCP of @AIP_Publishing
Preprint: https://arxiv.org/abs/2005.07046
A larger version of the poster can be found here: https://raw.githubusercontent.com/JaGeo/JaGeo.github.io/master/Poster_VCT_large.gif
A preprint of the publication is also already available: https://arxiv.org/abs/2005.07046
A longer thread with more detailed explanations will follow tomorrow.
A preprint of the publication is also already available: https://arxiv.org/abs/2005.07046
A longer thread with more detailed explanations will follow tomorrow.
I have made a video that illustrates the speedup reached with Gaussian Approximation Potentials.
It takes less than a minute with 8 threads to compute the phonon bandstructure of diamond-type Si after the potential is loaded.
It takes less than a minute with 8 threads to compute the phonon bandstructure of diamond-type Si after the potential is loaded.
We would like to speed up the calculation of phonon properties by using machine-learned interatomic potentials instead of DFT. This might enable us to discover new thermoelectric materials with low thermal conductivity in the future. #VCT2020Poster #VCT2020
Machine-learned interatomic potentials (we use Gaussian Approximation Potentials) learn energy and forces from reference data (most of the time DFT data). For more information on these potentials, see https://doi.org/10.1002/adma.201902765
We started our study by assessing a general-purpose potential for Si - GAP-18 ( http://doi.org/10.1103/PhysRevLett.104.136403). We calculated the phonon properties for a range of Si allotropes from the Materials Project ( http://materialsproject.org/ )
For diamond-type Si, the phonon bandstructures from GAP-18 and DFT are nearly indistinguishable. This is not surprising, the reference database for GAP-18 contains many diamond-type Si structures.
For clathrate-I, there are larger deviations between GAP-18 and DFT (more than 10% for some phonon frequencies). State-of-the-art DFT calculations typically show a mean relative error of -3.6% in comparison to experiment ( https://doi.org/10.1038/sdata.2018.65).
We also tested the error on the phonon bandstructure for 11 other dynamically-stable Si allotropes from the Materials Project. The phonon error was always larger than for diamond-type Si.
Our research question was now: Can we improve the accuracy of GAPs for phonons?
Our research question was now: Can we improve the accuracy of GAPs for phonons?
We then build a range of reference databases (including supercells with individual displacements, cells with random displacements, and bulk data from the initial GAP-18 database) and fitted several GAP models. We identified a strategy leading to excellent phonon accuracy of GAPs.
We arrived at a new potential that can arrive at excellent phonon properties for all 13 dynamically-stable Si allotropes. The phonon error is typically much lower than 0.3 THz!
The excellent properties of GAP-18 for liquid and amorphous silicon were kept in the new potential. We analysed this with the help of a melt-quench simulation. Along this simulation, we took snapshots where we compared the GAP results to DFT.
You can find the code to reproduce the study here: https://github.com/JaGeo/Phonons_GAPs_silicon
