Convective drying is an ubiquitous unit operation in chemical and allied industries; it is energy intensive and a significant contributor to the carbon footprint of a plant. Developing detailed models, a digital twin, of an industrial oven can enable rigorous energy optimization. In this context, we use a neural partial differential equation formalism to train a model of an industrial oven to capture the evolution of the moisture and temperature as a solid passes through. We also use an event function to capture the transition of the moisture and the temperature between the constant rate drying regime and the falling rate drying regime. We show that this hybrid model, even when trained on partially observed and spatially sparse industrial data, accurately captures the system dynamics and generalizes effectively to other spatial locations. We proffer that neural differential equations provide enough flexibility to model complex chemical processes and include domain knowledge to deal with limited and noisy data.
