Multispectral camouflage requires materials that simultaneously reduce detectability in microwave bands and infrared (IR) atmospheric windows. Achieving these objectives entails competing electromagnetic requirements and a high-dimensional design space, making conventional trial-and-error optimization inefficient. Here, a Bayesian optimization (BO)–driven inverse-design framework is reported for a lightweight metamaterial integrating broadband microwave absorption and low IR emissivity within a compact architecture composed of optically transparent constituents. The BO algorithm adaptively explores a high-dimensional design space spanning unit-cell geometric parameters and layer thicknesses, by iteratively updating a surrogate model using data from full-wave electromagnetic simulations. The optimized design achieves broadband microwave absorption exceeding 90% across 4.6–18 GHz (covering the C, X, and Ku bands) with a total thickness of 5.425 mm, approaching the Rozanov thickness bound. In the IR, the structure exhibits low band emissivity of 0.16 in the mid-wave IR (MWIR, 3–5 μ m) and 0.17 in the long-wave IR (LWIR, 8–14 μ m), leading to a reduced apparent radiation temperature and a shortened sensor lock-on range relative to a black-body reference.