Point defect formation at finite temperatures with machine learning force fields
Irea Mosquera-Lois, Johan Klarbring, Aron Walsh · 2024
Read Point defect formation at finite temperatures with machine learning force fields by Irea Mosquera-Lois, Johan Klarbring, Aron Walsh (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-16741.txt.
Introduction
Irea Mosquera-Lois, Johan Klarbring, and Aron Walsh study how heat changes the math of crystal defects. Point defects set the behavior of many working materials, but standard models treat the crystal as frozen, counting only internal energy. Real crystals vibrate, and atoms settle into shapes the frozen picture never sees. The team trained a machine learning force field to follow defect motion at real temperatures in cadmium telluride.
They weighed each flavor of entropy, from vibration to spin, and compared free energy methods from simple harmonic estimates to full thermodynamic integration. Shapes once dismissed as unstable turned out to be populated at room temperature. Heat raised the predicted count of one defect type a hundredfold, enough to change the material's expected behavior. The authors argue defect forecasts must include temperature.
Catalog record
| CREATOR | Irea Mosquera-Lois, Johan Klarbring, Aron Walsh |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-16741
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Point defect formation at finite temperatures with machine learning force fields Irea Mosquera-Lois, Johan Klarbring, Aron Walsh Abstract Point defects dictate the properties of many functional materials. The standard approach to modelling the thermodynamics of defects relies on a static description, where the change in Gibbs free energy is approximated by the internal energy. This approach has a low computational cost, but ignores contributions from atomic vibrations and structural configurations that can be accessed at finite temperatures. We train a machine learning force field (MLFF) to explore dynamic defect behaviour using Te i + 1 \mathrm{Te_{i}^{+1}} and V Te {}_{\text{Te}} +2 in CdTe as exemplars. We consider the different entropic contributions (e.g., electronic, spin, vibrational, orientational, and configurational) and compare methods to compute the defect free energies, ranging from a harmonic treatment to a fully anharmonic approach based on thermodynamic integration. We find that metastable configurations are populated at room temperature and thermal effects increase the predicted concentration of Te i + 1 \mathrm{Te_{i}^{+1}} by two orders of magnitude — and can thus significantly affect the predicted properties. Overall, our study underscores the importance of finite-temperature effects and the potential of MLFFs to model defect dynamics at both synthesis and device operating temperatures. I Introduction Point defects make or break material functionality [ 1 ] . They limit photovoltaic efficiency by acting as non-radiative recombination centres, control ionic conductivity in batteries, provide active sites for catalytic reactions, and platforms for quantum information technologies. Despite their profound effect on the macroscopic properties of crystals, they are present in dilute concentrations and thus render experimental characterisation challenging. As a result, a combination of experiment and theory is needed to understand defect behaviour. The key factor when modelling defects is their concentration, which is determined by the free energy of defect formation, g_{f} , at the synthesis or annealing temperature. Calculating g_{f} is however computationally challenging, and is thus typically approximated by the formation internal energy, u_{f}(0~\mathrm{K}) , i.e. g_{f}(T_{\rm synthesis})\approx u_{f}(0~\mathrm{K}) [ 2 , 3 ] . Inherent in this approximation is the assumption of a static framework, where most studies only consider the defect ground state structure at 0\text{\,}\mathrm{K} , and thus neglect metastable configurations that may be populated at the device operating temperature. Since the properties of a defect strongly depend on its geometry [ 4 , 5 , 6 , 7 , 8 , 9 ] , the predicted behaviour can be significantly affected when ignoring thermally accessible metastable configurations [ 10 , 11 , 12 , 13 , 14 , 15 , 16 ] . With the development of better computational resources, more accurate studies that go beyond this static 0\text{\,}\mathrm{K} approximation are becoming possible using ab-initio methods. In the last decades, many investigations have modelled entropic contributions for defects in elementary solids [ 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 , 59 , 60 , 61 ] . Thermal effects are harder to model in multinary semiconductors due to the higher number of possible intrinsic charged defects and required level of theory, but have been included for specific defects [ 62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 , 74 , 75 , 76 , 77 , 78 , 79 , 80 ] . However, most of these studies adopt several approximations: i) they only account for vibrational entropies, thereby neglecting other degrees of freedom (e.g., electronic, spin, orientational and configurational), and ii) they adopt the (quasi)harmonic approximation to model vibrational effects (e.g., assuming a quadratic potential energy surface for the interatomic bonds). The limitations of these approximations are system-dependent and not well investigated, and demonstrate the lack of a reliable and affordable approach to model thermal effects for defects. In this study, we target these limitations by considering all relevant entropic contributions and systematically comparing the different methods to calculate the defect formation free energy, ranging from a harmonic to a fully anharmonic approach based on thermodynamic integration (TI). To reduce the cost of these simulations, we use machine learning force fields as a surrogate model, which successfully map the defect energy surfaces ( Section II.1 ). We choose \mathrm{Te_{i}^{+1}} and V {}_{\text{Te}} +2 in CdTe as exemplar systems since they display potential energy surfaces of different complexity (e.g., presence versus lack of low-energy metastable configurations [ 12 ] ) ( Section II.2 ). By modelling the impact of thermal effects on their predicted defect concentrations, we find that these dominate when the defect undergoes symmetry-breaking structural reconstructions and has low-energy metastable configurations, thereby demonstrating the limitations of the idealised 0\text{\,}\mathrm{K} description ( Section II.3 ). II Results II.1 Machine learning force fields for defects Calculating the defect formation free energy, g_{f} , requires modelling the defect formation reaction by computing the free energy difference between products and reactants at the synthesis temperature. [ 3 , 2 ] For example, the formation of the positively charged tellurium interstitial, \mathrm{Te_{i}^{+1}} , can be described by the defect reaction \displaystyle\mathrm{Te}\xrightarrow{\mathrm{CdTe}}\mathrm{Te_{i}}^{+1}+e^{-}. (1) The corresponding formation free energy is defined from the sum of products (charged defect with an electron in the conduction band) minus the sum of reactants (pristine CdTe host and reservoir of Te). In the standard first-principles supercell formalism, this formation energy is given by \displaystyle\mathrm{g_{f}=g(Cd_{n}Te_{n+1}^{+1})-g(Te)-g([CdTe]_{n})}+E_{F}+E_{corr} (2) where Te represents the phase that acts as the external source of atoms during synthesis, E_{F} denotes the Fermi level and E_{corr} is the correction energy for charged defects. From these terms, the defect free energy, \mathrm{g(Cd_{n}Te_{n+1}^{+1})} , is the most challenging to compute due to the low symmetry and large supercells required to model defects (e.g., many force calculations in the (quasi)harmonic method). This is exacerbated when going beyond the harmonic approximation since computing the anharmonic free energy with TI requires many and long molecular dynamics runs. [ 81 , 82 ] To reduce the cost of free energy calculations, we employ machine learning force fields and train a separate MACE model [ 83 ] for each system involved in the defect formation reaction, targeting temperatures ranging from 100\text{\,}\mathrm{K} to the typical CdTe anneal temperature of 840\text{\,}\mathrm{K} [ 84 ] . All models show good accuracies with low mean and root mean square errors on the test set (see 1 and further discussion in Methods and Supporting Information (SI)). The accuracy of the defect models is further confirmed by mapping the one-dimensional path between the stable defect structures, which shows good agreement despite the small energy difference between the distinct configurations of \mathrm{Te_{i}^{+1}} ( 1 and SI S9 ). Table 1: Mean absolute errors and root mean square errors (shown in parentheses) of the test sets for energies, forces and stresses. The relatively high errors observed