Research map / Time series / Causal and generative time series
The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning
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 the Catechol benchmark, a time-series transient-flow dataset of 1227 solvent-mixture yield measurements for an allyl catechol rearrangement, and benchmarks GP, MLP, LLM-embedding, ODE, transfer, active learning and BO methods on it.
Problem
Yield prediction for continuous reaction conditions (temperature, residence time, solvent mixture) is poorly served by existing ML-ready chemistry datasets, which focus on discrete variables and are often inconsistent or not ML-ready.
Approach
Collect dense transient-flow time-series data where temperature, residence time and solvent ratios are varied continuously in a single run, and provide a benchmark with solvent featurizations, GP extensions, transfer learning from a related Ethyl dataset, and active learning/Bayesian optimization tasks.
Key results
- Spange-featurized GP achieves the best full-data NLPD (-5.663) and single-solvent NLPD (-5.793) among standard regressors, while MLP with Spange gets the lowest single-solvent MSE (0.010).
- LLM embeddings (RXNFP, ChemBERTa) perform poorly with frozen backbones (MSE 0.105 and 0.153 on full data).
- Transfer learning from the Ethyl dataset improves the Spange GP MSE from 0.030 to 0.020 on catechol data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Spange solvent descriptors instead of ignoring solvent | Gaussian process · Catechol single solvent | MSE: 0.014 -> 0.011 | Helped |
| Use Spange solvent descriptors instead of ignoring solvent | Gaussian process · Catechol full data | MSE: 0.011 -> 0.011 | No clear effect |
| Use Spange descriptors rather than RXNFP embeddings | MLP · Catechol full data | MSE: 0.105 -> 0.010 | Helped |
| Use Spange descriptors rather than ChemBERTa embeddings | MLP · Catechol full data | MSE: 0.153 -> 0.010 | Helped |
| Use a decomposed kernel (ablation) | GP (Spange) · Catechol single solvent | MSE: 0.011 -> 0.009 | Helped |
| Use a decomposed kernel (ablation) | GP (Spange) · Catechol full data | MSE: 0.011 -> 0.012 | Hurt |
| Model output correlations with a multitask GP (ablation) | GP (Spange) · Catechol full data | MSE: 0.011 -> 0.018 | Hurt |
| Apply input warping (ablation) | GP (Spange) · Catechol single solvent | MSE: 0.011 -> 0.011 | No clear effect |
| Transfer learning from the Ethyl dataset | GP (Spange) · Catechol product-yield prediction | MSE: 0.030 -> 0.020 | Helped |
| Transfer learning from the Ethyl dataset | MLP (Spange) · Catechol product-yield prediction | MSE: 0.027 -> 0.031 | Hurt |
| Use multi-objective Bayesian optimization instead of random sampling | Optimization benchmark · Multi-objective optimization, iteration 100 | GD: 0.054 -> 0.014 | Helped |
| Use multi-objective Bayesian optimization instead of random sampling | Optimization benchmark · Multi-objective optimization, iteration 100 | IGD: 0.092 -> 0.072 | Helped |
Surprising findings
- Many off-the-shelf algorithms fail to outperform a baseline GP that ignores solvent information entirely.
- Frozen LLM embeddings (RXNFP, ChemBERTa) give much worse regression than simple fixed featurizations.
Limitations
- Only two reactions are studied (allyl catechol and the Ethyl/allyl phenyl ether reaction).
- Only Gaussian processes are explored in depth; deep learning and Bayesian neural networks are left to future work.
- Acetic acid and acetonitrile experiments were excluded due to unintended side reactions.
Experimental setup
Datasets: Catechol benchmark (1227 solvent mixture points, 656 single solvent points) Allyl Phenyl Ether / Ethyl dataset (283 points)
Benchmarks: Catechol benchmark
Metrics: mean squared error negative log predictive density generational distance inverted generational distance maximum Pareto frontier error
Baselines: BaselineGP MLP NODE LODE Expected Improvement Upper Confidence Bound Random search
Builds on: Gaussian processes Thompson sampling ChemBERTa RXNFP
Closest papers by meaning
- CheMixHub: Datasets and Benchmarks for Chemical Mixture Property Prediction (NeurIPS 2025 · similarity 0.87)
- Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction (ICML 2025 · similarity 0.86)
- BOOM: Benchmarking Out-Of-distribution Molecular Property Predictions of Machine Learning Models (NeurIPS 2025 · similarity 0.86)
- Uncertainty-Aware Yield Prediction with Multimodal Molecular Features (AAAI 2024 · similarity 0.85)
- SC$^3$: A Multi-Solvent Solubility Challenge and Benchmark (NeurIPS 2026 · similarity 0.85)
- From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction (ICLR 2024 · similarity 0.84)
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.