Research map / Graph learning / GNN architectures
Geometric Mixture Models for Electrolyte Conductivity Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Curates CALiSol and DiffMix electrolyte conductivity datasets with 3D geometric graphs and proposes GeoMix, a Set-SE(3) equivariant model with a Geometric Interaction Network for intermolecular message passing, outperforming MLP, MolSets, and EGNN/TFN baselines.
Problem
Predicting ionic conductivity of multi-molecule electrolyte mixtures lacks standardized benchmarks and existing models do not capture 3D geometry and fine-grained intermolecular interactions while respecting mixture symmetries.
Approach
A geometry-aware mixture model that builds PCA-based local frames per molecule and learns node-pair transformation matrices and translations to pass equivariant messages across molecules, preserving Set-SE(3) equivariance; proportions are embedded into atom features.
Key results
- GeoMix-TFN reaches MSE 0.432 and Pearson r 0.987 on CALiSol, versus 1.107 for the best baseline TFN-linear.
- On DiffMix, GeoMix-TFN reaches MSE 0.035 and r 0.997, versus 0.195 for the best baseline EGNN-linear.
- In the temperature OOD split, GeoMix-TFN reaches MSE 2.354 versus 4.780 for TFN-linear.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GeoMix intermolecular geometric message passing to a TFN backbone | TFN-linear · CALiSol | MSE: 1.107 -> 0.432 | Helped |
| add GeoMix intermolecular geometric message passing to an EGNN backbone | EGNN-linear · DiffMix | MSE: 0.195 -> 0.088 | Helped |
| embed proportions in atomic features rather than apply scalar weighting to molecule embeddings (ablation) | GeoMix-EGNN · CALiSol | MSE: 3.657 -> 0.552 | Helped |
| allow an unconstrained transformation matrix rather than a unit-quaternion rotation (ablation) | GeoMix-EGNN · CALiSol | MSE: 0.702 -> 0.552 | Helped |
| allow node-pair transformations rather than one graph-pair transformation (ablation) | GeoMix-EGNN · CALiSol | MSE: 0.662 -> 0.552 | Helped |
| use attention-based readout rather than linear averaging (ablation) | GeoMix-EGNN · CALiSol | MSE: 0.851 -> 0.552 | Helped |
| add Noisy Nodes loss (ablation) | GeoMix-EGNN · CALiSol | MSE: 1.213 -> 0.552 | Helped |
| use deterministic strict PCA instead of PCA (ablation) | GeoMix-EGNN · CALiSol | MSE: 0.552 -> 0.628 | Hurt |
Surprising findings
- GeoMix-EGNN degrades sharply under 0.3 Å coordinate noise on the TFN-based model (MSE 4.517 for GeoMix-TFN vs 0.631 for GeoMix-EGNN), indicating TFN-based models are sensitive to coordinate perturbation.
- Strictly equivariant deterministic PCA frames performed slightly worse than standard PCA.
Limitations
- Datasets are limited in scope; no MD-based microscopic structural information and no sodium- or potassium-ion electrolytes.
- The encoder architecture is relatively simple.
- Standard PCA frames are not strictly equivariant, and the strict variant performs slightly worse.
Experimental setup
Datasets: CALiSol DiffMix PubChem conformations RDKit conformers Materials Project structures
Benchmarks: CALiSol DiffMix
Metrics: MSE Pearson r MAE Spearman r
Baselines: MLP MM-MoLFormer MolSets-Conv MolSets-SAGE EGNN-att TFN-att EGNN-linear TFN-linear
Builds on: EGNN TFN Noisy Nodes MolSets DiffMix
Closest papers by meaning
- CheMixHub: Datasets and Benchmarks for Chemical Mixture Property Prediction (NeurIPS 2025 · similarity 0.87)
- ELoRA: Low-Rank Adaptation for Equivariant GNNs (ICML 2025 · similarity 0.83)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.83)
- Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.83)
- HAGO-Net: Hierarchical Geometric Message Passing for Molecular Representation Learning (AAAI 2024 · similarity 0.83)
- Learning the Electronic Hamiltonian of Large Atomic Structures (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.