Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Nada Alghamdi, Paolo de Angelis, Pietro Asinari, Eliodoro Chiavazzo

TL;DR

This work develops a multi-layer screening pipeline that combines MACE foundational ML force fields with DFT+U validation to identify novel cathode materials from the Energy-GNoME database. By validating against experimental cathodes, the authors demonstrate that MACE-r2scan accurately predicts average intercalation voltages and gravimetric energy densities with substantial speed advantages over full DFT, and that MACE-OMAT efficiently filters dynamically unstable candidates via phonon analysis. The screening workflow reduces an initial pool of 615 candidates to a focused set of feasible materials across Li/Na/Mg/K/Ca chemistries, with 43 candidates advancing to DFT+U verification and a final shortlist of promising targets. The study also shows a good correlation between ML predictions, Energy-GNoME figures of merit, and DFT+U results, supporting Energy-GNoME as a viable pre-screening resource for rapid cathode discovery and pointing to future work on transport-property predictions via molecular dynamics.

Abstract

The development of new battery materials, particularly novel cathode chemistries, is essential for enabling next generation energy storage technologies. In this work, we set up a screening procedure on the Energy-GNoME database for identifying novel cathode candidates. We use MACE foundational models as a first layer of screening, we assess dynamical stability and estimate the average voltage and gravimetric energy density. Following that, we apply physically motivated reasoning to identify the most promising candidates. Furthermore, we refine the average voltage prediction of selected promising candidates using DFT+U and provide the list of selected materials using this protocol. This work delivers two key outcomes: validation of the foundational MACE models high-throughput screening approach and suggestions for cathode candidates for the development of next-generation batteries. Finally, a fair comparison between the MACE predictions and the readily available figures of merit reported in the Energy GNoME database is demonstrated on the examined materials.

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

TL;DR

This work develops a multi-layer screening pipeline that combines MACE foundational ML force fields with DFT+U validation to identify novel cathode materials from the Energy-GNoME database. By validating against experimental cathodes, the authors demonstrate that MACE-r2scan accurately predicts average intercalation voltages and gravimetric energy densities with substantial speed advantages over full DFT, and that MACE-OMAT efficiently filters dynamically unstable candidates via phonon analysis. The screening workflow reduces an initial pool of 615 candidates to a focused set of feasible materials across Li/Na/Mg/K/Ca chemistries, with 43 candidates advancing to DFT+U verification and a final shortlist of promising targets. The study also shows a good correlation between ML predictions, Energy-GNoME figures of merit, and DFT+U results, supporting Energy-GNoME as a viable pre-screening resource for rapid cathode discovery and pointing to future work on transport-property predictions via molecular dynamics.

Abstract

The development of new battery materials, particularly novel cathode chemistries, is essential for enabling next generation energy storage technologies. In this work, we set up a screening procedure on the Energy-GNoME database for identifying novel cathode candidates. We use MACE foundational models as a first layer of screening, we assess dynamical stability and estimate the average voltage and gravimetric energy density. Following that, we apply physically motivated reasoning to identify the most promising candidates. Furthermore, we refine the average voltage prediction of selected promising candidates using DFT+U and provide the list of selected materials using this protocol. This work delivers two key outcomes: validation of the foundational MACE models high-throughput screening approach and suggestions for cathode candidates for the development of next-generation batteries. Finally, a fair comparison between the MACE predictions and the readily available figures of merit reported in the Energy GNoME database is demonstrated on the examined materials.

Paper Structure

This paper contains 18 sections, 8 equations, 8 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Results
  3. Validation of computational methods
  4. Average voltage and gravimetric energy density
  5. Phonon based dynamical stability
  6. Screening Energy-GNoME database
  7. MACE screening
  8. DFT screening
  9. Methods comparison
  10. Method
  11. Convex hull and voltage profile
  12. Phonon-based dynamical stability
  13. Computational framework
  14. Technical setup
  15. Conclusions
  16. ...and 3 more sections

Figures (8)

  • Figure 1: Illustration of the convex hull and average voltage of LiCoO2. Experimental data is from HU202361. Additional supporting figures are provided in the Supplementary Information.
  • Figure 2: Comparison of the average voltages and gravimetric energy densities obtained by MACE-r2scan, DFT-r2scan, and experiment. Experimental data were computed from the voltage profiles: LiCoO2 HU202361, LiFePO4ROWDEN202197, Li2MnO3Rana2014Structural, NaCoO2reddy2015high, KVPO4FK-ion-cathode, MgV2O4Hu2020High
  • Figure 3: MACE-OMAT achieved a 95% accuracy in predicting stable structures and a 54% accuracy for unstable ones. Based on the phonon database in Ref. loew2025universal, which contains 7,360 stable and 1,585 unstable structures.
  • Figure 4: Overview of the screening protocl followed in this study
  • Figure 5: Parity plot comparing the average voltage predicted by MACE-r2scan against predictions from Energy-GNoME
  • ...and 3 more figures