Screening novel cathode materials from the Energy-GNoME database using MACE machine learning force field and DFT

The development of new battery materials, particularly novel cathode chemistries, is essential for enabling next generation energy storage technologies. In this work, we employ a multi-fidelity screening protocol combining the Energy-GNoME confident criteria, foundational MACE machine-learning force fields (MLFF), and physically motivated heuristic filters to identify novel intercalation cathodes for post-lithium batteries, namely: Na-, K-, Mg-, and Ca-ion batteries. Foundational MACE models are used to efficiently asses dynamical stability, thermodynamical stability, average voltage, and theoretical specific energy, enabling a rapid screening of candidates. For the most promising cathodes, voltage predictions are refined using DFT+U calculations. This work delivers three key outcomes: i) establishing and validating a robust high-throughput screening approach for cathode materials with foundational MLFF models; ii) suggestions for cathode candidates for the development of next-generation of batteries; iii) a fair comparison between the MACE predictions and the readily available figures of merit reported in the Energy-GNoME database on the examined materials.

💡 Research Summary

In this work the authors present a multi‑fidelity high‑throughput screening workflow that combines the Energy‑GNoME database, foundational MACE machine‑learning force fields (MLFF), and density‑functional theory with Hubbard‑U corrections (DFT+U) to identify promising cathode materials for post‑lithium batteries (Na, K, Mg, and Ca ion chemistries). The pipeline begins by selecting only those entries in Energy‑GNoME with an “AI‑expert” confidence score above 90 %, reducing the initial pool of 20 454 candidates to 615 structures. These structures are then evaluated with two pre‑trained MACE models (MACE‑OMAT and MACE‑r2SCAN), which provide rapid force calculations, phonon spectra, formation energies, average intercalation voltages, and theoretical specific energies.

Filter 1 uses the MACE‑derived phonon dispersions to discard dynamically unstable compounds (negative phonon modes). Filter 2 computes equilibrium voltage profiles and specific energies; any material with an average voltage outside the 2.1–6 V window or a specific energy below 250 Wh kg⁻¹ is eliminated. To further improve synthetic feasibility, two heuristic filters are applied: Filter 3 removes structures belonging to space groups that are rarely observed in experimentally reported inorganic crystals (only the 24 most common space groups are retained), while Filter 4 excludes compounds containing radioactive, highly toxic, or scarce elements, favoring those composed of abundant, low‑cost constituents.

After these four successive filters, only a few dozen candidates remain. For this final set, the authors perform DFT+U calculations to refine the voltage profiles, using a Hubbard U term appropriate for transition‑metal d‑states. This step provides a higher‑accuracy benchmark while keeping computational costs manageable because of the prior massive reduction in candidate numbers.

The methodology is validated on six experimentally characterized cathodes: three lithium‑based (LiCoO₂, LiFePO₄, Li₂MnO₃) and three non‑lithium systems (NaCoO₂, KVPO₄F, MgV₂O₄). MACE‑r2SCAN reproduces experimental average voltages within ±0.1 V and yields specific energies consistent with measured values. Notably, for MgV₂O₄ and KVPO₄F, DFT‑r2SCAN underestimates voltages due to the use of non‑spin‑polarized calculations (a limitation of the Quantum ESPRESSO version employed), whereas MACE, trained on a diverse dataset that includes various magnetic configurations, maintains accurate predictions.

A direct comparison between the MACE‑derived metrics and those originally reported in the Energy‑GNoME database reveals systematic deviations, indicating that the database’s purely ML‑based predictions can be overly optimistic or pessimistic for certain chemistries, especially where training data are sparse (e.g., Ca‑based cathodes). The authors argue that incorporating physics‑based MLFF evaluations can substantially improve the reliability of large‑scale materials databases.

Overall, the study demonstrates that a combined MLFF‑DFT workflow can reduce the search space for new cathode materials by four orders of magnitude, identify a tractable set of promising Na‑, K‑, Mg‑, and Ca‑ion cathodes, and provide a validated, reproducible protocol for future high‑throughput battery materials discovery.