Dye molecules are widely applicable in areas ranging from biological imaging and diagnostics to dye-sensitized solar cells. To guide chemical design in these applications, there has been a decades-long effort to better predict the optical properties of dyes using physics-based or statistical models. However, these studies were often limited in their scope or representation ability; they have typically focused on a single dye family, restricted themselves to a single solvent, or used a fixed set of fingerprints and descriptors. In this work, we leverage recently published datasets of experimental data along with time-dependent density functional theory (TD-DFT) calculations to train a message-passing neural network (MPNN). Our MPNN predicts the highest UV/Vis absorption peak location and full width at half maximum (FWHM) for a given dye-solvent pair. We demonstrate that our automatically-learned, optimized representation of the dye-solvent pair outperforms traditional fingerprint and descriptor representations on a wide range of dye families. Additionally, we find that including results from TD-DFT as features in our model improves predictive ability over the model or TD-DFT on their own. Our best model achieves a mean absolute error of less than 10nm on a held-out test set for both tasks of predicting the location of the highest absorption peak and the corresponding FWHM. This predictive ability will accelerate the design of dye molecules when coupled with generative models and experimental synthesis and characterization.

Predicting absorption spectra of molecular dyes using deep learning

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.