Empirical and physics-based computational approaches to explosive power, brisance, and underwater detonation response: A critical review with selected machine-learning applications

Reliable interpretation of energetic-material performance requires methods that connect thermochemical properties and detonation behavior with clearly defined experimental endpoints. This critical narrative review examines empirical correlations, thermochemical calculations, numerical models, and selected machine-learning (ML) applications relevant to three domains: explosive power, brisance, and underwater detonation response. The principal emphasis is on established empirical and physics-based computational approaches used to estimate intrinsic detonation properties or interpret test-derived performance measures. Empirical correlations are transparent and computationally efficient, but their reliability depends on the data and conditions for which they were developed. Thermochemical and numerical models represent physical processes more explicitly, although their predictions remain sensitive to input data, model assumptions, and boundary conditions. ML studies are included where they directly inform the reviewed domains, including QSPR/read-across estimation of thermochemical properties, prediction of density and detonation descriptors, and surrogate modeling of underwater fluid or structural response. These studies address different targets and use labels derived from experiments, empirical equations, or numerical simulations; consequently, their reported error metrics are not directly comparable. Representative methods are compared in terms of prediction target, required inputs, assumptions, validation evidence, computational burden, and limitations. The review is not intended as a comprehensive survey of ML architectures or general-purpose energetic-material discovery, because the available evidence is uneven across the three performance domains. Instead, it identifies where empirical, physics-based, and data-driven methods can complement one another and where further validation is needed. A proposed integration strategy is outlined that combines empirical baselines, thermochemical or numerical calculations, ML surrogates, physics-informed constraints where governing equations are available, active learning, and calibrated uncertainty estimates. This strategy is presented as a research direction rather than an independently validated framework spanning all three domains.