Energetic materials are solids that can release significant energy upon stimuli. They can constitute propellants, explosives, fuels, and pyrotechnics used in aerospace, mining, and defense industries. To reduce the sensitivity of the materials and enhance safety, energy materials are often manufactured as two-phase composites, with a softer binder as the host matrix that holds the explosive crystals, such as HMX (High-melting Explosive) in place. This design, referred to as Polymer-bonded explosives (PBX), enables the molding, shaping, and uniformity of the materials, leading to improved predictable performance. Nevertheless, the characterization of the energy localization often requires a material model capable of handling extremely large deformation of phase transformation. This talk reports recent progress on forward and inverse problems for modeling the HMX and PBX enabled by machine learning. We attempt to create mathematical models of HMX expressed in symbolic form. To avoid the difficulty of training the Kolmogorov-Arnold network, we introduce an alternative technique to learn neural additive basis in projected feature space to control the expressivity-speed trade-off.