Motivation
High-throughput computational screening of cathode materials is bottlenecked by the computational cost of Density Functional Theory (DFT) calculations. A single DFT relaxation can require 10–100 CPU hours, making exhaustive screening of large candidate spaces prohibitively expensive.
Furthermore, the majority of randomly sampled candidates are thermodynamically unstable (Ehull > 50 meV/atom), resulting in wasted computational resources. Machine learning pre-screening offers an efficient filtering mechanism to prioritize candidates before expensive DFT validation.
Approach
We employ MACE-MP-0 (Multi-Atomic Cluster Expansion), a state-of-the-art equivariant foundation model for materials science pretrained on the Materials Project (Batatia et al., 2023). MACE uses higher-order equivariant message passing with multi-body interactions, enabling highly accurate prediction of energy above the convex hull (Ehull).
Model Configuration (v1-Li-Cathode)
Architecture
- • MACE-MP-0 "medium" backbone (~3.5M params)
- • 5-member deep ensemble (seeds 42–46)
- • Frozen backbone, fine-tuned last interaction block
- • Custom regression head: quantile outputs (q10, q50, q90) + stability classification (p_stable, p_metastable)
Training Details
- • Dataset: 17,227 Transition Metal Oxides (TMOs)
- • Splitting: SOAP-LOCO (Leave-One-Cluster-Out)
- • Target: Ehull in eV/atom
- • Post-hoc symmetric conformal calibration (90% coverage)
Uncertainty Quantification
We combine three complementary sources of uncertainty following best practices in deep ensemble methodology (Lakshminarayanan et al., 2017):
# Aleatoric (data noise) — per-model quantile regression
σ_aleatoric = mean(q90 − q10) across ensemble
# Epistemic (model ignorance) — inter-model disagreement
σ_epistemic = std(q50 across 5 members)
# Total uncertainty
σ_total = √(σ_aleatoric² + σ_epistemic²)
Prediction intervals are calibrated via symmetric conformal prediction on the validation set, guaranteeing 90% coverage. A conformal delta is added to raw quantile intervals to achieve valid coverage under distribution shift (Vovk et al., 2005).
Validation Methodology
We use SOAP-LOCO (Smooth Overlap of Atomic Positions — Leave One Cluster Out) splitting to evaluate generalization to unseen chemical families. This approach clusters materials by structural similarity and holds out entire clusters during validation, providing a rigorous test of extrapolation capability.
Performance Results
Governance: APPROVED (6/6 checks passed) — Ranking (Spearman > 0.5), calibration (90% coverage), KEEP precision (> 85%), false-kill rate (< 2%), and decision-making all verified.
0.663
Spearman ρ
Test set ranking
0.030
MAE
eV/atom (test)
92.7%
KEEP Precision
123 / 1,013 KEEP'd
0.0%
False Kill
Rate
Known limitation: LOCO (leave-one-cluster-out) performance degrades significantly (Spearman ≈ 0, coverage 72%). The model is reliable for in-distribution cathodes but should not be trusted for structurally novel polymorphs.
Note: Primary metrics computed on test split (1,013 materials). Model scope: Transition Metal Oxide cathodes from the Materials Project.
Decision Policy
Materials are classified into actionable tiers using conformally-calibrated quantile predictions from the 5-member ensemble:
| Action | Criterion | Interpretation |
|---|---|---|
| KEEP | q90 < 0.05 eV ∧ p_stable > 0.8 | High confidence stable → Prioritize for DFT |
| KEEP | q90 < 0.10 eV ∧ p_stable > 0.7 | Likely metastable → Worth DFT validation |
| KILL | q10 > 0.10 eV | Confident unstable → Skip DFT |
| MAYBE | Otherwise | Uncertain → Manual review recommended |
References
- Batatia, I., Benber, P., Chiang, B., et al. (2023). A foundation model for atomistic simulation.arXiv:2401.00096.
- Deng, B., Zhong, P., Jun, K., Riebesell, J., Han, K., Bartel, C. J., & Ceder, G. (2023). CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5(9), 1031–1041.
- Jain, A., Ong, S. P., Hautier, G., Chen, W., Richards, W. D., Dacek, S., ... & Persson, K. A. (2013). Commentary: The Materials Project: A materials genome approach to accelerating materials innovation.APL Materials, 1(1), 011002.
- Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles.Advances in Neural Information Processing Systems, 30.
- Bartók, A. P., Kondor, R., & Csányi, G. (2013). On representing chemical environments.Physical Review B, 87(18), 184115.
- Vovk, V., Gammerman, A., & Shafer, G. (2005). Algorithmic Learning in a Random World.Springer. (Conformal prediction)
Programmatic Access
CathodeScreen exposes a RESTful API for batch predictions. Upload CIF files and receive JSON responses with predicted Ehull, uncertainty estimates, and decision recommendations.
{
"material_id": "mp-1234567",
"formula": "LiCoO2",
"ehull_pred_q50": 0.023,
"ehull_pred_q10": 0.012,
"ehull_pred_q90": 0.034,
"sigma_epistemic": 0.005,
"sigma_total": 0.008,
"p_stable": 0.91,
"decision": "KEEP",
"conformal_coverage": 0.90
}