Tropical weather and climate variability consists of numerous circulation and cloud systems that occur on a hierarchy of temporal and spatial scales ranging from the scale of the convective cloud of 1km to 10 km to planetary intraseasonal disturbances such as the Madden Julian oscillation (MJO) and monsoon dynamics. These systems interact nontrivially with each other and with the global dynamics. Unfortunately, despite many recent improvements in computing power and in our understanding of the earth system’s physics, state-of-the-art climate models still simulate very poorly such cloud systems and the associated rainfall distributions, especially in the tropics. This is believed to be due to the underlying subgrid models used in the state-of-the-art climate models, which are deterministic in nature as they are based on a quasi-equilibrium theory which makes convection slaved to the slow-resolved scale dynamics. As such the climate modes fail to represent the often chaotic variability due to the interactions, with each other and with the environment, of the various cloud types that characterize tropic cloud systems. In this talk, we discuss a new promising direction to this effect using a stochastic lattice interacting particle system to emulate the bottom-up organization of cloud systems over multiple temporal and spatial scales. The lattice system allows the representation of the main cloud types the associated interactions, using observations as a guiding principle and to machine learn some key driving parameters of the model. This leads to a better representation of the aforementioned tropical modes of variability including the MJO and monsoon dynamics.