Speaker
Description
Machine-learning (ML) interatomic potentials trained with first-principles data promise steep advances for the predictive-quality modeling and simulation of molecules and materials. At a computational cost that is significantly reduced compared to direct first-principles calculations, such ML potentials allow to address larger system sizes or perform more extensive dynamical simulations and sampling. While typically not as cost efficient as classical force fields with a fixed functional form, they straightforwardly include reactivity and, most importantly, can seamlessly be improved by additional training data.
This versatility also has its downsides though. With the ML potential itself completely void of any physics, the training data need to adequately cover the structural and chemical space of interest. Depending on the application, the underlying multiple first-principles calculations for the training data could then themselves start to become a computational bottleneck. The latter is, e.g., particularly pronounced for surface science or interfacial applications such as heterogeneous catalysis or batteries. There, training structures may necessarily extend to large supercell calculations, which even on an efficient semi-local density functional theory (DFT) level may constitute a formidable computational burden. This calls for data-efficient training protocols that establish a reliable ML potential with a minimum number of DFT training data (of tractable system sizes).
Achieving such an efficient training is particularly challenging if the targeted chemical space is not known a priori at the beginning of a study – as is notably the case during global surface structure determination (SSD). To this end, we present a general and data-efficient iterative training protocol that blends the creation of new training data with the actual surface exploration process [1]. Demonstrating this protocol with the SSD of low-index facets of rutile IrO2 and RuO2, the involved simulated annealing on the basis of the refining GAP readily identifies a number of unknown terminations even in the restricted sub-space of (1 x 1) surface unit cells. The metal-rich terminations are thermodynamically predicted to become most stable in reducing environments and are subsequently indeed identified experimentally [2]. Reminiscent of complexions discussed in the context of ceramic battery materials, they are completely unexpected for rutile oxides. Correspondingly, they escaped the traditional approach to SSD in form of testing a set of candidate structures devised by the researcher.
[1] J. Timmermann et al., J. Chem. Phys. 155, 244107 (2021).
[2] J. Timmermann et al., Phys. Rev. Lett. 125, 206101 (2020).
| Abstract Number (department-wise) | TH 18 |
|---|---|
| Department | TH (Reuter) |
