We present a hybrid physics–AI digital-twin framework for Earth-system modelling that tightly integrates machine learning with first-principles climate dynamics [1,2]. The approach embeds a physics-constrained neural network (PCNN) within a host general circulation model (GCM), enabling data-driven representation of unresolved processes while enforcing physical consistency. We introduce the core architectural components of the PCNN-GCM, including CondensNet, a condensation-aware AI surrogate designed to model cloud microphysics within the dynamical core [3]. By construction, the framework addresses two central challenges in hybrid climate modelling: maintaining physical fidelity and ensuring long-term numerical stability. Physical constraints are embedded directly into the learning architecture, yielding bounded, interpretable, and conservation-consistent predictions, while preserving the stability of multi-decadal climate integrations. The resulting hybrid model substantially improves robustness and interpretability compared to purely data-driven parameterizations, without sacrificing computational efficiency. This work establishes a principled pathway toward trustworthy, scalable hybrid climate models, advancing digital-twin capabilities for simulating and predicting complex atmospheric processes under current and future climate conditions.