Fine-tuning a MACE foundation model fixes sodium clustering in hard carbon, cutting formation-energy error from 32.43 to 1.65 meV/atom
Synopsis
Pretrained MACE atomistic foundation models describe sodium intercalation in graphite-like domains but treat the metallic attraction between sodium atoms on graphene as repulsion, so they miss the clustering and pore filling that shape hard-carbon storage. Targeted fine-tuning on 1,985 sodium-cluster configurations plus 300 expanded graphite and sodium reference configurations lowers the formation-energy mean absolute error on the sodium-cluster data from 32.43 to 1.65 meV/atom and raises R² from -36.32 to 0.87. After fine-tuning on a broader sodium-carbon set, the model still gives a 2.03 meV/atom formation-energy error on sodium-filled carbon nanotubes excluded from training, and it links closed-pore radius to plateau voltage.
Interpretation
Pretrained MACE-family atomistic foundation models miss the clustering of sodium into metallic aggregates, the interaction that decides whether sodium stays dispersed or aggregates across hard-carbon local motifs. These models had already shown near-density-functional-theory (DFT) accuracy for amorphous carbon and sodium intercalation in graphite, and were expected to serve as cheap surrogates for DFT in battery modeling. On a dataset of sodium clusters on graphene, the pretrained MATPES model reaches a mean absolute error of 32.43 meV/atom with R² of -36.32; for two sodium atoms on graphene, DFT shows a local energy minimum where the model predicts repulsion.
Targeted fine-tuning on a small set of sodium-cluster and reference configurations restores DFT-level sodium-carbon energetics and the correct ordering of cluster stability. Before fine-tuning, the model deviated strongly on sodium-cluster formation energies and failed to reproduce how graphite total energy changes with interlayer distance. On the graphene sodium-cluster data, the mean absolute error falls from 32.43 to 1.65 meV/atom and R² rises from -36.32 to 0.87; compact clusters are correctly more stable than dispersed sodium, and graphite AA/AB reference-state errors drop to a few meV/atom.
Extending the fine-tuning data to defective graphene, intercalated graphite, pseudographitic carbon, amorphous sodium-carbon structures and expanded references keeps accuracy on sodium-carbon configurations outside the training set. Fine-tuning only on the small cluster-and-reference set gave larger errors on the broader configurations and missed chemistries such as amorphous structures and adsorption at monovacancies. On a randomly split test set, formation-energy mean absolute error is 10.70 meV/atom with R² near 1.00 and force-component error is 0.023 eV/Å with R² of 0.99, nearly matching the training set; on sodium-filled carbon nanotubes excluded from the dataset, formation-energy error is 2.03 meV/atom with R² of 0.97.
Thermodynamic analysis with the fine-tuned model ties closed-pore radius to the plateau voltage of sodium storage, with small pores contributing at higher voltage. Pore filling had often been treated as corresponding to a single plateau voltage. For idealized cylindrical closed pores with radii from 0.59 to 0.91 nm, convex-hull analysis gives one stable structure per pore, and larger radii store more sodium but activate at lower voltage; the model and DFT differ by about 1 mV in voltage for CNT(15,0) and CNT(16,0), and by more for CNT(17,0).
Perspective
The demonstrated scope covers sodium-carbon motifs modeled as graphene, defective graphene, graphite, pseudographitic carbon, amorphous sodium-carbon structures and idealized cylindrical closed pores; energies in the fine-tuning dataset were computed without van der Waals dispersion, whereas the sodium-filled carbon nanotube and closed-pore voltage analyses included D3-BJ dispersion corrections. Closed pores are approximated by single-walled carbon nanotubes, and the authors set the intended application range to atomic force components below 10 eV/Å.
The authors report that the broader-dataset model has low R² on adsorption-type configurations, with formation energies following the DFT trend but shifted overall, partly attributable to the graphite AB and sodium BCC reference states; on sodium-filled carbon nanotubes the model and DFT predict different ground-state structures, and the force-component R² is negative, which the authors attribute to DFT reference forces being close to zero with small variance.
