Abstract The discovery of organic photocatalysts is fundamentally limited by the vastness of chemical space and the scarcity of standardized experimental data. Conventional graph neural networks often fail to navigate this landscape, particularly at “activity cliffs” where structural isomers with identical 2D topologies but distinct 3D electronic environments exhibit drastically different catalytic performance. Here, we introduce a Fourier-based Kolmogorov-Arnold Graph Attention Network (Fourier-KAGAT), a geometrically enhanced architecture that integrates learnable Fourier-based activation functions into the graph message-passing step, enabling the model to capture subtle stereoelectronic effects that fixed non-linearities miss. Fourier-KAGAT achieves a 70% recall on a blind test set of conjugated molecules, significantly outperforming random forest baselines (30%), and systematically resolves 20 of 28 activity cliff pairs between constitutional isomers where standard GNNs fail entirely. Furthermore, the model’s latent representations correlate strongly with four DFT-calculated electronic properties governing the photocatalytic mechanism—electron affinity ( R 2 = 0.78), optical gap ( R 2 = 0.66), singlet-triplet gap ( R 2 = 0.36), and exciton binding energy ( R 2 = 0.51)—consistently outperforming both MLP-GAT and random forest baselines, confirming that Fourier-KAGAT learns representations encoding multiple physicochemical properties rather than exploiting dataset biases. This approach offers a data-efficient pathway for the computational design of scalable catalytic materials.
Peivaste et al. (Sat,) studied this question.