GenGLem: a generative framework for capturing chemical cliffs in energetic materials

High-throughput prediction of explosive properties is currently bottlenecked by the prohibitive computational cost of potential energy surface (PES) sampling and the inability to account for complex environmental conditions. Here, we develop GenGLem, a dual-pathway generative framework featuring latent space exploration and microstructure guidance. GenGLem fundamentally transforms configurational sampling from traditional trajectory-based search into direct, end-to-end latent space generation. By integrating unconditional latent exploration and conditional diffusion pathways during exploration pipeline, the framework incorporates thermodynamic variables and environmental factors - including temperature, pressure, and solvent - to enable targeted, environment-induced sampling. Validation demonstrates that 71.7–91.8% of generated conformers are energetically favorable, exhibiting superior stability and diversity compared to simulated annealing with a ~ 1200-fold speedup. Critically, GenGLem captures pronounced environment-induced structural reorganization, maintaining high consistency with experimental observations. This study establishes a new framework bridging deep learning and molecular dynamics for rapid PES sampling, offering a transformative tool for advancing molecular property prediction and computational chemistry applications.