High-Throughput-Screening Workflow for Predicting Volume Changes by Ion Intercalation in Battery Materials

Abstract

Mechanical stresses and strains developing locally within the microstructure of active ion-battery-electrode materials during charge-discharge cycles can compromise their long-term stability. In this context, crystalline compounds exhibiting low volume changes are of particular interest. Atomistic simulations can be employed to quantify the volume change of the crystal structure upon intercalation and deintercalation of ions and to elucidate the local mechanisms underlying the global structural response. While density functional theory (DFT) offers a robust and accurate framework for such calculations, its computational cost limits its applicability for large-scale screening of diverse intercalation structures and sites. In this work, we present a workflow designed to prioritize candidate materials for subsequent detailed characterization. The workflow calculates the volume change upon intercalation using atomic-level features and a machine-learning model for bond-length prediction. The bond-length predictions are based on the assumption that bonds between the same ionic species in similar local coordination environments exhibit comparable lengths across different crystallographic structures. The model was trained on a DFT-generated dataset, which inherently defines the chemical space in which reliable predictions can be expected. We demonstrate the workflow's utility by screening approximately 1,175,000 transition-metal oxides and fluorides, followed by DFT validation of the most promising candidates. The proposed workflow enables filtering of large candidate sets and accelerates the potential discovery of low volume change intercalation materials for batteries.

Figures (6)

• Figure 1: Schematic representation of the intercalation process: In a solid-solution mechanism, the intercalation of an atom (green sphere) into a host structure alters the local structure and can influence the global volume (i.e., the volume of the unit cell), but does not drastically modify the topology of the host lattice. The zoom-in illustrates a possible attraction of the surrounding atoms to the newly inserted atom.
• Figure 2: Workflow to predict the volume change upon intercalation of ions into a host structure ($S^{\mathrm{host}}_\mathrm{DFT}$). In the first step, this results in a structure $S^{n=0}_\mathrm{ML}$ with the unchanged volume $V^{\mathrm{host}}_\mathrm{DFT}$. For this structure, atomic and bond feature vectors are created to predict the bond lengths ($l^{\mathrm{pred.}}_{ij}$). By systematically shifting atomic positions and cell vectors, the structure is then iteratively adjusted to minimize the differences between the actual bond lengths $l^{\mathrm{actual}}_{ij}$ and the predicted values $l^{\mathrm{pred.}}_{ij}$ based on the model $\mathcal{M}_{\mathrm{Bond}}$. The iteration is continued until self-consistency is achieved, yielding the final structure $S^{\mathrm{final}}_\mathrm{ML}$ with volume $V^{\mathrm{final}}_\mathrm{ML}$. The procedures highlighted in light blue and red are described in detail in Sec. \ref{['sec:theory:lsop']} and \ref{['sec:mod:theory:vol']}, respectively.
• Figure 3: LSOP structure motifs used in this work. The motifs describe the coordination of the central atom (orange) by the neighboring atoms (dark blue). The values 90$^\circ$, 104.5$^\circ$, 120$^\circ$ and 150$^\circ$ for $\alpha$ were used for the bent bond and the values 90$^\circ$ and 120$^\circ$ for $\alpha$ were used for the see-saw shaped coordination leading to 25 LSOPs in total. The coordination number, by which the similarity to a motif is multiplied, is given below each motif and corresponds to the number of dark blue neighboring atoms. The Figure was adapted from Fig. 6 in [N. E. R. Zimmermann and A. Jain, RSC Adv. 2020, 10, 6063--6081]. Available under a CC-BY 3.0 license. Copyright N. E. R. Zimmermann and A. Jain
• Figure 4: Left: The DFT-derived bond lengths between two atoms $i$ and $j$ plotted against the sum of their ionic radii. Right: The DFT-derived bond lengths plotted against the obtained bond lengths using the ML model $\mathcal{M}_{\mathrm{Bond}}$. The MAE and the Pearson correlation coefficient ($r$) for both datasets are given in the respective plots. Each data point is shown with an opacity of 0.01, so that 100 overlapping points result in full coverage.
• Figure 5: Precision–-recall curve for the model $\mathcal{M}_{\mathrm{Vol.}}$, evaluated using bond lengths predicted by the model $\mathcal{M}_{\mathrm{Bond}}$. Selected thresholds of $|\Delta V_{\mathrm{ML}}|$ are indicated with circles. A baseline corresponding to random selection of structure pairs is shown.