Research map / Probabilistic and scientific modeling / Molecules and proteins
CheMixHub: Datasets and Benchmarks for Chemical Mixture Property 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.
Introduces CheMixHub, a curated benchmark of 11 chemical-mixture property prediction tasks from 7 datasets, with four generalization splits and baseline deep-set/attention and XGBoost models, plus an Arrhenius physics head.
Problem
Chemical mixture property prediction lacks standardized public datasets and benchmarks, and existing ML evaluations use inconsistent splits that do not test generalization across mixture sizes, molecules, and temperatures.
Approach
Consolidates 11 mixture tasks from 7 public sources with standardized mole-fraction conversion, four splitting strategies (random, mixture size, leave-molecules-out, temperature), and a benchmark of permutation-invariant deep set/attention models with GNN, MolT5, and RDKit features, plus an Arrhenius-form output head for temperature-dependent properties.
Key results
- On 5-fold random CV, XGBoost with MolT5 or RDKit features is competitive with deep models, e.g., best MAE on drug solubility ln(S) 0.028 and IlThermo ln(kappa) 0.071 (MolT5 XGB).
- Adding an Arrhenius physics head to the best deep model reduces temperature-split MAE on NIST ln(eta) from 0.079 to 0.048 and raises Pearson rho to 0.980.
- Fine-tuning from the Miscible Solvents density task to Delta H_mix or Delta H_vap transfers poorly (rho 0.348 and 0.644), while the reverse direction transfers well (rho 0.955 and 0.929).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use an Arrhenius physics-based predictive head instead of a regular fully connected head (ablation) | Best deep learning model · IlThermo ln(κ), temperature-range exclusion | MAE: 0.354 ± 0.152 -> 0.284 ± 0.025 | Helped |
| Use an Arrhenius physics-based predictive head instead of a regular fully connected head (ablation) | Best deep learning model · IlThermo ln(η), temperature-range exclusion | MAE: 0.162 ± 0.096 -> 0.127 ± 0.011 | Helped |
| Use an Arrhenius physics-based predictive head instead of a regular fully connected head (ablation) | Best deep learning model · NIST ln(η), temperature-range exclusion | MAE: 0.079 ± 0.006 -> 0.048 ± 0.013 | Helped |
| Use an Arrhenius physics-based predictive head instead of a regular fully connected head (ablation) | Best deep learning model · Solid Polymer Electrolytes ln(κ), temperature-range exclusion | MAE: 0.481 ± 0.228 -> 0.363 ± 0.015 | Helped |
| Use an Arrhenius physics-based predictive head instead of a regular fully connected head (ablation) | Best deep learning model · Temperature-range exclusion splits across four temperature-dependent tasks | Pearson ρ: 0.923 ± 0.084 -> 0.987 ± 0.002; 0.968 ± 0.033 -> 0.988 ± 0.003; 0.946 ± 0.006 -> 0.980 ± 0.011; 0.941 ± 0.024 -> 0.961 ± | Helped |
| Use XGBoost on a temperature-range exclusion split rather than a random split | XGBoost · Temperature-dependent tasks | MAE | Hurt |
| Use a mixture interaction module, either attention or DeepSets | Deep learning mixture models · CheMixHub random-split tasks | MAE | Mixed |
| Fine-tune a model trained on ΔH_vap for density prediction | Miscible Solvents · Miscible Solvents ρ | MAE: 0.003 ± 0.000 -> 0.026 ± 0.002 | Hurt |
| Fine-tune a model trained on ΔH_mix for density prediction | Miscible Solvents · Miscible Solvents ρ | MAE: 0.003 ± 0.000 -> 0.021 ± 0.001 | Hurt |
| Fine-tune a model trained on MS-ΔH_vap for MON prediction | Fuel mixtures · MON | MAE: 4.570 ± 0.348 -> 33.199 ± 1.606 | Hurt |
| Fine-tune a model trained on MON for MS-ΔH_vap prediction | Miscible Solvents · Miscible Solvents ΔH_vap | MAE: 0.071 ± 0.002 -> 1.582 ± 0.095 | Hurt |
| Use a model trained on NIST-full for zero-shot prediction on NIST viscosity | Viscosity prediction models · NIST ln(η) | MAE: 0.030 ± 0.001 -> 6.806 ± 0.012 | Hurt |
Surprising findings
- XGBoost, a strong baseline on random splits, degrades markedly on temperature-range exclusion splits (e.g., IlThermo ln(kappa) MAE rising to 0.386).
- Density-trained models transfer poorly to harder nonlinear properties like Delta H_mix and Delta H_vap, while harder-task models transfer well to density.
Limitations
- Polymers are reduced to monomeric units, which limits representation of periodic polymer structure.
- GNNs were not pre-trained, and 3D conformer information was not studied.
- Cross-dataset transfer and zero-shot results are weak outside tasks with similar value ranges.
Experimental setup
Datasets: Miscible Solvents IlThermo NIST viscosity (TRC) Drug solubility Solid polymer electrolytes Olfactory mixtures Fuel mixtures (MON)
Benchmarks: CheMixHub 11 tasks
Metrics: MAE Pearson correlation Kendall tau
Baselines: XGBoost with RDKit descriptors XGBoost with MolT5 embeddings DeepSets self-attention set model GNN-based molecular representation
Builds on: DeepSets Set Transformer MolT5 XGBoost Bilodeau et al. NIST viscosity curation
Closest papers by meaning
- ChemPile: A 250 GB Diverse and Curated Dataset for Chemical Foundation Models (NeurIPS 2025 · similarity 0.88)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.87)
- Geometric Mixture Models for Electrolyte Conductivity Prediction (NeurIPS 2025 · similarity 0.87)
- The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning (NeurIPS 2025 · similarity 0.87)
- Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials (NeurIPS 2026 · similarity 0.86)
- BOOM: Benchmarking Out-Of-distribution Molecular Property Predictions of Machine Learning Models (NeurIPS 2025 · similarity 0.86)
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.