Automation of machine learning driven interatomic potential generation for predicting vibrational properties Christina Ertural 1 , Volker L. Deringer 2 , Janine George 1,3 1 Federal Institute for Materials, Research and Testing, Department Materials Chemistry, Unter den Eichen 87, 12205 Berlin, Germany. 2 Department of Chemistry, Inorganic Chemistry Laboratory, University of Oxford, Oxford OX1 3QR, UK 3 Friedrich Schiller University Jena, Institute of Condensed Matter Theory and Solid-State Optics, Max-Wien-Platy 1, 07743 Jena, Germany. *e-mail: christina.ertural@bam.de Stability and thermal conductivity behaviour of materials are dependent on their vibrational properties, e.g., low thermal conductivity is beneficial for thermoelectric properties. 1–3 Employing the quasi-harmonic approximation (QHA) to investigate the phononic properties of a compound in the established way, i.e., density functional theory- based methods, 4 takes many calculation steps and consumes a lot of computational resources to arrive at the desired results. Even then one may encounter situations where the QHA reaches its limits and calculations have to be extended by software packages like TDEP 5 or hiPhive 6 for temperature dependent phonon calculations or higher-order force constants. In a recent work, 7 some of us demonstrated that using machine learning (ML) driven interatomic potentials (e.g., Gaussian approximation potential, 8 GAP) opens up a beneficial alternative route to the traditional computation way of phonons. In this work, we automate the generation of such machine- learned interatomic potentials (MLIP) in a Python code-based workflow. This workflow is based on the automation tools atomate2 9 and pymatgen 10 and combines the automatic DFT computations with the automated fitting of MLIP (starting with GAP). Automation facilitates the testing of different strategies for reference data generation as well as different hyperparameters for the MLIP fitting procedure for single elements as well as more complex compounds. In addition, the potentials might be automatically validated for their purpose, as, for example discussed in Ref. 11. We plan to present the automation and the first potentials derived and validated based on the automation and their suitability for phonon computations with the aim to provide the resulting potentials for storage in databases. References 1. Snyder, G. J.; Toberer, E. S. Nat. Mater. 2008 , 7 (2), 105–114.
2. He, J. et al.Adv. Funct. Mater. 2022 , 32 (14), 2108532. 3. Wehmeyer, G. et al.Appl. Phys. Rev. 2017 , 4 (4), 041304. 4. Togo, A.; Tanaka, I. Scr. Mater. 2015 , 108 , 1–5. 5. Hellman, O. et al.Phys. Rev. B 2013 , 87 (10), 104111. 6. Eriksson, F.; Fransson, E.; Erhart, P. Adv. Theory Simul. 2019 , 2 (5), 1800184. 7. George, J. et al.J. Chem. Phys. 2020 , 153 (4), 044104. 8. Bartók, A. P. et al.Phys. Rev. Lett. 2010 , 104 (13), 136403. 9. Mathew, K. et al.Comput. Mater. Sci. 2017 , 139 , 140–152. 10. Ong, S. P. et al.Comput. Mater. Sci. 2013 , 68 , 314–319. 11. Morrow, J. D.; Gardner, J. L. A.; Deringer, V. L. J. Chem. Phys. 2023 , 158 (12), 121501.
P17DD
Made with FlippingBook Learn more on our blog