Bridged azobenzene derivatives are important photo-responsive molecular switches with wide-ranging applications in optoelectronics and information storage, owing to their reversible isomerization around the –N=N– bond. However, elucidating the microscopic mechanisms underlying their isomerization pathways remains challenging, as conventional spectroscopic approaches often struggle to resolve complex intramolecular coupling effects and to establish robust structure–spectrum correlations. Here, we present an integrated, autonomous research workflow that combines robotic experimentation, infrared (IR) and Raman spectroscopy, quantum chemical simulation, and machine learning (ML) for structural analysis of bridged azobenzenes. Central to our approach is the attention-based 1D convolutional neural network (ATT-CNN), which quantitatively predicts the critical C–N=N–C dihedral angle directly from vibrational spectra, achieving a correlation coefficient of r = 0.99 and a mean absolute error of just 5°. By leveraging pre-trained models and transfer learning, our method demonstrates strong generalizability across different chemical environments and experimental conditions. The attention mechanism within ATT-CNN further enables holistic, interpretable evaluation of vibrational features, ensuring reliable structure determination even in the presence of routine experimental variations. Orchestrated by the ChemAgents multi-agent platform, the workflow tightly integrates experimental design, automated computational and experimental data acquisition, and machine learning analysis in a closed-loop, data-driven manner. This strategy not only advances fundamental understanding of azobenzene isomerization mechanisms, but also establishes a generalizable blueprint for intelligent, automated laboratory research in dynamic molecular systems.
Shen et al. (Tue,) studied this question.