Industrial drying and granulation plants that run fluidized beds are significant consumers of natural gas, and many production sites operate several such lines side by side rather than a single unit. Waste heat leaving one line could in principle preheat the intake air of another, but whether this pays off depends on temperature levels as much as on quantities: much of the recoverable heat is released at low temperature, where it cannot replace gas firing on the same line. Exploiting it therefore requires coordination across processes and across the site. At the same time, production plants run at fixed, product-specific settings, so their operating records contain little variation to learn from, and safety requirements rule out letting an experimental controller act on a running plant.

TwinLink builds a digital twin of such a multi-line facility, using fluidized bed spray granulation as the example process. A physics-based process model is combined with time-series machine learning architectures (recurrent networks, Transformers and state-space models) that predict energy consumption, product moisture and exhaust-air temperature from the history of measurable operating conditions. 
Once validated, the models serve as a fast stand-in for the plant, letting a reinforcement learning agent explore operating strategies across the site without running the physics simulation, or the plant itself, for every candidate decision. The objective is to reduce energy consumption while keeping product quality within specification.