Bond Energy Descriptors Enable Machine Learning With Limited Data: Design of Capping Agents for Thermoplastic Polyurethane Recycling

Big‐data approaches to chemistry falter when experimental data remain scarce, expensive, and noisy. Here we demonstrate that physics‐informed descriptors overcome this limitation and enable accurate predictions from severely limited datasets. Taking capped diisocyanate deblocking as a test case, we introduce descriptors derived from the bond energies from bond orders and populations (BEBOP) method that capture electronic structure and bonding at low computational cost. Trained on only 19 experimental compounds, BEBOP‐informed LASSO regression predicts deblocking temperatures within an RMSE of 11 °C under fully nested leave‐one‐out cross‐validation over a 227–323 °C range, while conventional molecular descriptors fail to produce any meaningful correlation. We validate the resulting physically interpretable model on two new compounds, one within the training range and one outside of it. Embedding physical insight into descriptors rather than relying on algorithmic complexity reduces data requirements by an order of magnitude and offers a generalizable strategy for molecular design problems wherever large datasets are impractical.