Machine learning prediction of energy band gaps in curcumin derivatives

Machine learning (ML) regression models using SMILES-encoded molecular fingerprints were developed to predict the band gaps of 19 synthesized curcumin derivatives. The goal was to develop a ML-based framework capable of estimating the band-gap values prior to synthesis, thereby guiding chemists in their experimental effort. Training data for the model were obtained from the Harvard Clean Energy Project (HOPV-15), restricted to 52,000 non-fullerene acceptor (NFA) organic molecules with density functional theory (DFT)-calculated band gaps. Multiple fingerprinting schemes, including Morgan fingerprints, MACCS keys, atom-pairs, and FCFP, were used to train a pool of regression models. These models were then applied to the 19 curcumin derivatives for band-gap prediction and compared to DFT simulation calculations. The accuracy of the predictions was enhanced by using the Most Common Substructure (MCS) approach, narrowing it to molecules sharing key curcumin motifs. This refinement reduced the average predictive error by 30% compared with using the full HOPV-15 dataset, highlighting the benefit of targeted feature selection. Finally, the model was applied to virtually generate and evaluate 1,225 unsynthesized curcumin derivatives by permuting side groups on the BF2-curcuminoid backbone, identifying ten candidates with predicted band gaps above 3.2 eV, which were selected as a proof-of-concept case study to demonstrate the capability of the proposed machine learning framework for targeted virtual screening. This workflow demonstrates how ML-guided screening can uncover promising curcumin-based chromophores prior to synthesis, reducing experimental costs while providing chemical insights into structure–property relationships.