What if your machine learning model could ‘see’ chemistry like a chemist? The secret isn’t more data - it’s smarter descriptors.

What if your machine learning model could ‘see’ chemistry like a chemist? The secret isn’t more data - it’s smarter descriptors.

‍

Machine learning doesn’t understand chemistry by default. It needs a way to represent it.

In reaction optimization, that often comes from chemical descriptors. 

Instead of treating solvents and ligands as different categories, descriptors encode their physical, steric, and electronic properties in a form that models can learn from.

At ReactWise, we use solvent descriptors to capture effects like polarity and hydrogen-bonding effects, and ligand descriptors to represent steric bulk and electronic properties such as orbital energies.

This allows optimization to move beyond memorizing conditions. Models can interpolate between related ligands or solvents, allowing the transfer of insight across campaigns

For example, in a cross-coupling reaction, increasing ligand steric bulk can accelerate reductive elimination, while stronger electron donation can favour oxidative addition. 

When these features are encoded explicitly, models can show more than just “ligand A works” - they can highlight which descriptors/molecular properties are driving the desired selectivity, or conversion.

The effects can be visualized through a variety of plots - such as waterfall or beeswarm - that illustrate the key drivers of performance.

Descriptors help open the black box of machine learning by making data-driven optimization more interpretable, transferable, and reliable.

‍