Scientific Machine Learning (SciML) aims to develop models that can understand complex physical phenomena described by Partial Differential Equations (PDEs), a task that becomes particularly challenging when dealing with multiphase and multiphysics systems like boiling processes in heat transfer. This paper introduces Decoupled Mamba (DMamba), a novel model based on the Mamba3D architecture, designed for spatio-temporal volumetric data. DMamba features a grid-independent patch embedding layer, decoupled encoders, a bottleneck layer, and a unified decoder.The proposed model overcomes limitations of existing approaches, such as Fourier Neural Operator (FNO) and UNet-based models, which struggle to accurately capture local, discontinuous features in multiphase systems. By employing a linear-complexity operator, DMamba reduces parameter count while retaining key model benefits, leading to a more efficient and accurate solution for modeling complex multiphase and multiphysics behaviors.Our results demonstrate that DMamba achieves state-of-the-art performance in predicting temperatures and velocities across a variety of boiling scenarios, outperforming Unet, FNO variants, and recent transformer-based models in predictive accuracy. Additionally, DMamba exhibits robust generalization capabilities across diverse boiling scenarios, illustrating its potential to significantly advance the modeling of intricate multiphase systems. The source code is available at https://anonymous.4open.science/r/bubble-ml-6546.