Slide 1

Slide 1 text

MRS Fall 2024 virtualatoms.org Discovering highly anisotropic dielectric crystals Alex Ganose Imperial College London virtualatoms.org

Slide 2

Slide 2 text

MRS Fall 2024 virtualatoms.org Dielectrics have wide technological applications Light emitting diodes Capacitors Liquid crystal displays Photovoltaics DRAM Micro-electronics

Slide 3

Slide 3 text

MRS Fall 2024 virtualatoms.org Dielectric anisotropy enables novel function Birefringence LCDs, polarisers, medical imaging Emitters Polarised LEDs and lasers Emerging applications Direct dark matter detectors

Slide 4

Slide 4 text

MRS Fall 2024 virtualatoms.org Symmetry controls structure of dielectric tensors

Slide 5

Slide 5 text

MRS Fall 2024 virtualatoms.org Symmetry controls structure of dielectric tensors

Slide 6

Slide 6 text

MRS Fall 2024 virtualatoms.org Materials Project dielectric dataset 6,706 structures with dielectric tensors from Materials Project

Slide 7

Slide 7 text

MRS Fall 2024 virtualatoms.org https://e3nn.org Equivariant GNN enforces symmetry of matter Tensor product message passing rotate

Slide 8

Slide 8 text

MRS Fall 2024 virtualatoms.org Model ensures correct symmetry of tensors

Slide 9

Slide 9 text

MRS Fall 2024 virtualatoms.org Scalar dielectric prediction competitive with SOTA Model Test MAE coGN 0.309 MegNet 0.339 DimeNet++ 0.340 matbench-dielectric We predict full tensor then average, other models predict average directly

Slide 10

Slide 10 text

MRS Fall 2024 virtualatoms.org Equivariance provides marginal improvement

Slide 11

Slide 11 text

MRS Fall 2024 virtualatoms.org Anisotropy ratio skewed towards 1 Percentage error ~ 6% OK performance given ar is not directly trained

Slide 12

Slide 12 text

MRS Fall 2024 virtualatoms.org Equivariance essential for good performance

Slide 13

Slide 13 text

MRS Fall 2024 virtualatoms.org Screening the Materials Project for high anisotropy 0D 1D 2D 3D 1 2 3 4 Predicted Anisotropy Ratio 0 2 4 6 8 Anisotropy Ratio 0 10 20 30 Percent Training data Anisotropy favoured in 1D and 2D crystals (obviously)

Slide 14

Slide 14 text

MRS Fall 2024 virtualatoms.org HT DFT typically run by materials databases

Slide 15

Slide 15 text

MRS Fall 2024 virtualatoms.org The Materials Project registered users crystalline compounds molecular compounds band structures elastic tensors dielectric tensors 250k 150k 31k 52k 13k 7k

Slide 16

Slide 16 text

MRS Fall 2024 virtualatoms.org MP website is tip of the iceberg Slide credit: Matthew Horton

Slide 17

Slide 17 text

MRS Fall 2024 virtualatoms.org MP website is tip of the iceberg MP Software Stack Slide credit: Matthew Horton

Slide 18

Slide 18 text

MRS Fall 2024 virtualatoms.org Goal is to make HT possible for everyone what is the PBE elastic tensor of GaAs? results researcher workflows ├─ band structure ├─ surface energies ├─ elastic tensor ├─ raman spectrum └─ free energy workflow engine

Slide 19

Slide 19 text

MRS Fall 2024 virtualatoms.org Atomate2 makes this process easy run many different properties of many materials results database researcher

Slide 20

Slide 20 text

MRS Fall 2024 virtualatoms.org │ ├─ point defects ├─ magnetic ordering ├─ electron-phonon coupling ├─ lobster analysis ├─ electron mobility ├─ ab initio molecular dynamics ├─ ferroelectric properties ├─ electrode properties ├─ surface absorption ├─ thermal expansion ├─ lattice thermal conductivity └─ ...and more │ single-point ─┤ geometry optimisation ─┤ band structure ─┤ density of states ─┤ spin-orbit coupling ─┤ hybrid functionals ─┤ SCAN functionals ─┤ elastic tensor ─┤ dielectric tensor ─┤ piezoelectric tensor ─┤ harmonic phonons ─┤ free energy ─┘ Workflow library atomate2

Slide 21

Slide 21 text

MRS Fall 2024 virtualatoms.org structure, dir_name structure, dir_name dir_name, structure structure dir_name, structure stress, structure dir_name, structure dir_name, structure dir_name dir_name output output raw normalmode_frequencies, structure, 'born', normalmode_eigenvecs epsilon_ionic, epsilon_static epsilon_static 'frequency' 'ibands' 'dir_name' dir_name, structure dir_name output 'dir_name' output hse tight relax 1 hse tight relax 2 hse static dense uniform generate_elastic_deformations run_elastic_deformations fit_elastic_tensor dielectric bulk static deformation generate_wavefunction_coefficients amset calculate_polar_phonon_frequency sum calculate_deformation_potentials run_amset_deformations Workflows translate into a series of low-level tasks Example: electronic transport with VASP and AMSET

Slide 22

Slide 22 text

MRS Fall 2024 virtualatoms.org Supported DFT packages Many workflows are “code agnostic” and support multiple DFT software packages + MLFFs e.g. mace-mp-0, ORB, eqV2, SevenNet…

Slide 23

Slide 23 text

MRS Fall 2024 virtualatoms.org Atomate2 standardises knowledge All past and present knowledge, from many research groups, everyone previously in those groups, and their collaborators, about how to run calculations Slide credit: Anubhav Jain

Slide 24

Slide 24 text

MRS Fall 2024 virtualatoms.org Try atomate2 today Installation pip install atomate2 Code github.com/ materialsproject/atomate2 Support matsci.org/atomate

Slide 25

Slide 25 text

MRS Fall 2024 virtualatoms.org Our model successfully identifies highly anisotropic crystals Selected 137 materials with predicted ar greater than 2.5 Validated with density- functional perturbation theory through atomate2 0D 1D 2D 3D 1 2 3 4 Predicted Anisotropy Ratio

Slide 26

Slide 26 text

MRS Fall 2024 virtualatoms.org Our model successfully identifies highly anisotropic crystals 0 2 4 6 8 Anisotropy Ratio 0 10 20 30 Percent Training data Selected 137 materials with predicted ar greater than 2.5 Validated with density- functional perturbation theory through atomate2

Slide 27

Slide 27 text

MRS Fall 2024 virtualatoms.org Our model successfully identifies highly anisotropic crystals 0 2 4 6 8 Anisotropy Ratio 0 10 20 30 Percent Training data New discovery Selected 137 materials with predicted ar greater than 2.5 Validated with density- functional perturbation theory through atomate2

Slide 28

Slide 28 text

MRS Fall 2024 virtualatoms.org New materials identification

Slide 29

Slide 29 text

MRS Fall 2024 virtualatoms.org Conclusions Equivariant models vital for accurate tensorial predictions on small datasets Atomate2 aims to make high-throughput easy by integrating with common DFT packages Standardising workflows accelerates your research and with potential for generative model validation

Slide 30

Slide 30 text

MRS Fall 2024 virtualatoms.org Acknowledgements Anubhav Jain Andrew Rosen Janine George Janosh Riebesell Guido Petretto David Waroquiers Gian-Marco Rignanese Jimmy-Xuan Shen Nick Winner Aaron Kaplan Thomas Purcell Aakash Ashok Naik Christina Ertural Matt McDermott Matthew Evans Mingjian Wen Matthew Horton …and many more Atomate2 & jobflow contributors Dielectrics Leo Lou Funding EPSRC Compute ARCHER2 Imperial MMMHub

Slide 31

Slide 31 text

MRS Fall 2024 virtualatoms.org Acknowledgements Anubhav Jain Andrew Rosen Janine George Janosh Riebesell Guido Petretto David Waroquiers Gian-Marco Rignanese Jimmy-Xuan Shen Nick Winner Aaron Kaplan Thomas Purcell Aakash Ashok Naik Christina Ertural Matt McDermott Matthew Evans Mingjian Wen Matthew Horton …and many more Atomate2 & jobflow contributors Dielectrics Leo Lou Funding EPSRC Compute ARCHER2 Imperial MMMHub And you for your attention!