[Paper Review] Machine Learning for Electrode Materials: Property Prediction via Composition

Hao Wu, Cameron Hargreaves+2 more|arXiv (Cornell University)|Mar 8, 2026

Machine Learning in Materials Science0 citations

TL;DR

The paper benchmarks three composition-based ML models (MODNet, CrabNet, RF@Magpie) for predicting electrode-material properties from composition using the Materials Project Battery dataset, finding CrabNet most consistently accurate across metrics and validation schemes.

ABSTRACT

In this work, we benchmark three leading Machine Learning (ML) frameworks-MODNet, CrabNet, and a random forest model based on Magpie feature-for predicting properties of battery electrode materials using the Materials Project Battery Explorer dataset. We evaluate these models based on predictive accuracy, visualize numerical features using two-dimensional embeddings, and quantify performance using standard metrics. Our results demonstrate that CrabNet consistently outperforms the other models across all tests. To validate these findings, we employ robust statistical methods: bootstrap resampling and two cross-validation (CV) strategies (leave one cluster out and stratified 5-fold CV), comparing each model against a control baseline. In addition, we apply unsupervised clustering on MODNet-derived features using t-SNE and DBSCAN, revealing coherent material groupings without prior labels. This analysis confirms the robustness of the evaluated models and underscores the potential of ML-driven approaches for accelerating the electrode materials discovery. However, our study also identifies practical limitations and quantifies challenges associated with integrating ML models into materials science workflows. Despite these constraints, our findings suggest that ML models are highly effective for early-stage compositional screening in the battery industry. This work provides a foundation for future research on ML applications in materials discovery.

Motivation & Objective

• Assess the predictive performance of composition-based ML models on battery electrode properties.
• Compare MODNet, CrabNet, and RF@Magpie using a common feature set (Magpie features).
• Evaluate models with robust validation schemes including LOCO and stratified CV.
• Visualize high-dimensional features via 2D embeddings and assess clustering and representativeness of materials.
• Provide a benchmark and insights for composition-level screening in electrode discovery.

Proposed method

• Featurize electrode compositions using Magpie-based features (input vectors: MODNet 273 features, CrabNet 199, Magpie 21).
• Train three models: MODNet (neural network with feature selection via Normalized Mutual Information), CrabNet (transformer-like attention network on mat2vec embeddings), and RF@Magpie (random forest with Magpie features).
• Evaluate prediction accuracy on gravimetric capacity, volumetric capacity, and average voltage using mean absolute error (MAE) and SMAE (normalized MAE).
• Perform bootstrap resampling and two CV strategies: leave-one-cluster-out (LOCO) and stratified 5-fold CV, with clusters derived from DBSCAN on MODNet features.
• Use t-SNE/UMAP for 2D embeddings and DBSCAN for clustering to analyze structure and chemical similarity of materials.
• Compare results to a mean-predict baseline and report robustness across working ions and dataset size.

Figure 2 : Distribution of working ions in the electrode materials dataset.

Experimental results

Research questions

• RQ1 How do composition-based ML models (MODNet, CrabNet, RF@Magpie) perform in predicting electrode properties from composition alone?
• RQ2 Does CrabNet consistently outperform other models across multiple properties and validation schemes?
• RQ3 What is the impact of dataset size and class distribution (working ions) on model accuracy?
• RQ4 How do 2D embeddings and clustering relate to material groups and predictive performance?
• RQ5 What are practical limitations of applying ML to composition-level electrode discovery?

Key findings

• CrabNet consistently achieves the best predictive accuracy across gravimetric capacity, volumetric capacity, and average voltage under 5-fold CV and LOCO/Stratified CV.
• RF@Magpie generally underperforms relative to MODNet and CrabNet in MAE/SMAE and exhibits higher errors, especially under cross-validation schemes that test generalization.
• On gravimetric capacity, CrabNet achieves MAE values around 24.730 (18.126 in the 2σ-filtered data) and SMAE around 0.284 (0.208), with strong R2 near 0.84; similar robustness is observed for volumetric capacity and voltage.
• Cluster analysis (DBSCAN) on embedded MODNet features identifies 14 clusters, with chemical representativeness via ElMD mean representative materials; clustering aligns with known chemistries (e.g., Li-based clusters include LFP).
• Bootstrap analysis shows predictive errors decrease as dataset size increases, underscoring the value of larger compositional datasets for ML in materials discovery.
• LOCO CV yields higher errors than stratified CV, indicating model robustness varies with out-of-distribution tests; CrabNet maintains superior performance across schemes.

Figure 5 : 2D map of t-SNE embeddings of the materials using input features from MODNet. The points have been colored based on DBSCAN clustering. A total of 14 clusters are identified. The representative material from each cluster, as selected by ElMD mean representative, is indicated together with