Metal-organic frameworks (MOFs) demonstrate complex behaviours under changes in temperature, mechanical stress, or guest adsorption. These thermodynamic stimuli often induce reversible or irreversible transitions to states that lack long-range or framework order. Yet, these disordered states are functional and stand out due to their structural adaptability, e.g., the capacity to dissipate strain without significant material degradation. However, both experimental and computational challenges in analysing these highly disordered atomic structures leave the transition behaviours of such MOFs – is the transition (ir)reversible and does it retain some order? – unpredictable, preventing their rational design and application. Herein, I will address this gap by developing a multiscale in silico approach based on unified machine learning potentials (MLPs) to elucidate the mechanisms behind these order-to-disorder transitions and predict which MOFs transition along a given transition pathway. I will build such a detailed understanding by examining how systematically varying the constituent building blocks in two archetypal MOF platforms impacts their transition behaviour. This challenge requires me to overcome the traditional confines of crystalline modelling by developing MLPs and coarse-grained potentials for disordered structures and complement them by experimental validation, with the overall goal of enhancing our collective insight into these intriguing transition phenomena.