Cross-linking mass spectrometry (XL-MS) is a powerful tool for probing protein structures. While conventional chemical cross-linkers react with specific residues with defined chemistry, photo-cross-linkers, despite their superior reactivity, have been hindered by incomplete mechanistic understanding and lack of analytical framework. Here, we show that diazirine-based photo-cross-links are inherently MS-cleavable, generating composite backbone and side-chain fragments that complicates spectral interpretation. Yet, leveraging the side-chain fragmentation fingerprints, we developed a machine learning model that, when integrated with existing search algorithms, significantly improves ion coverage and reduces false discovery rate. Furthermore, we engineered a homo-bifunctional diazirine cross-linker, enabling true on-demand photo-cross-linking. Applying this workflow, we captured transient tetrameric assemblies of human Hsp90β and revealed structural transitions in association equilibrium upon heat stress, features inaccessible with conventional chemical cross-linking. Together, our work establishes a new paradigm in XL-MS, combining temporal resolution of photo-activation with analytical confidence for residue-level structural insights.