catechol_solvent_selection

Repository for the code and data on the catechol dataset for solvent selection and machine learning.

Installation

The requirements for this project can be installed using:

pip install -e .

or

pip install -r requirements.txt --no-deps

We recommend installing these in a fresh virtual environment, with Python version 3.11 or greater. We use uv to manage dependencies, but you don't need to install this yourself. If you wish to add more packages to the requirements, and you don't have uv installed, create a PR/issue.

Dataset

Here, we provide a brief overview of the dataset in this repo.

The dataset is divided into two:

• catechol_single_solvent_yields contains only the single-solvent data
• catechol_full_data_yields contains the full data set with mixture solvents

We also provide some pre-computed featurizations, which can be looked up with the SOLVENT NAME column.

Single solvent columns

Below is a table of all the columns in the catechol_single_solvent_yields csv:

Name	Type	Description
EXP NUM	int	Experiment index; all rows with the same EXP NUM will use the same solvent
Residence Time	float	Time (in minutes) of the reaction
Temperature	float	Temperature (in Celsius) of the reaction
SM	float	Quantity of starting material measured (yield %)
Product 2	float	Quantity of product 2 measured (yield %)
Product 3	float	Quantity of product 3 measured (yield %)
SOLVENT_NAME	str	Chemical name of the solvent; used as a key when looking up featurizations
SOLVENT_RATIO	list[float]	Ratio of component solvents [1]
{...} SMILES	str	SMILES string representation of a molecule

[1] This is different than the ratios in the solvent ramp experiments. Here, a single solvent has two component molecules, eg. the solvent "Acetonitrile.Acetic Acid" has two compounds. The SOLVENT_RATIO gives the ratio between these compounds. Most solvents consist of only a single compound, so the ratio will be [1.0].

Inputs: Residence Time, Temperature, SOLVENT NAME

Outputs: SM, Product 2, Product 3

Full data columns

The full data contains some additional columns, since these experiments ramp between two solvents:

Name	Type	Description
SolventB%	float	Percent concentration of solvent B; the rest of the solvent is made up of solvent A
SOLVENT {A/B} NAME	str	Chemical name of the solvents; used as a key when looking up featurizations

Inputs: Residence Time, Temperature, SOLVENT A NAME, SOLVENT B NAME, SolventB%

Outputs: SM, Product 2, Product 3

Running experiments

All of the scripts used to carry out the experiments are in scripts/, including

• eval_solvent_ramps.py
• eval_single_solvents.py
• eval_transfer_learning.py
• eval_active_learning.py
• eval_bayes_opt.py

Each of these files takes, as argument, a model and a featurization, as well as an additional configuration string. See the script README, as well as the individual scripts, for further details.

Referencing

If you use this datatset, please cite us as below:

@InProceedings{boyne2025catechol,
  title     = {The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning},
  booktitle = {Advances in Neural Information Processing Systems},
  author    = {Boyne, Toby and Campos, Juan S. and Langdon, Becky D. and Qing, Jixiang and Xie, Yilin and Zhang, Shiqiang and Tsay, Calvin and Misener, Ruth and Davies, Daniel W. and Jelfs, Kim E. and Boyall, Sarah and Dixon, Thomas M. and Schrecker, Linden and Folch, Jose Pablo},
  year      = {2025},
}