Cross-Modal\nPrediction of Spectral and Structural\nDescriptors via a Pretrained Model Enhanced with Chemical Insights

Proposing\nand utilizing machine learning descriptors\nfor chemical\nproperty prediction and material screening have become a cutting-edge\nfield in artificial intelligence-enabled chemical research. However,\na single descriptor typically captures only partial features of a\nchemical object, resulting in an information deficiency and limiting\ngeneralizability. Obtaining a comprehensive set of descriptors is\nessential but challenging, especially when accessing some microlevel\nstructural and electronic features due to technological limitations.\nHerein, we exploit multimodal chemical descriptors to construct an\nencoder–decoder machine learning framework that enables the\ncross-modal prediction of spectral and structural descriptors. By\npretraining the model to endow it with chemical insights, the multimodal\ndata fusion is implemented in a descriptor-encoded hidden layer. The\nmodel’s capabilities are validated in the system of CO/NO adsorption\non Au/Ag surfaces, demonstrating successful reciprocal prediction\nof infrared spectra, Raman spectra, and internal coordinates. This\nwork provides a proof-of-concept for the feasibility of cross-modal\npredictions between different chemical features and will significantly\nreduce the machine learning model’s dependence on complete\nphysicochemical parameters and improve its multitarget prediction\ncapabilities.