Topology-Aware Graph Neural Networks for Multistage Spray Drying

Spray drying is one of the most widely used processes for converting liquids into powders, used in the production of chemicals, pharmaceuticals, and many food products. Yet operating multistage spray dryers reliably remains difficult: the process couples turbulent airflow, heat and mass transfer with phase change, droplet collisions, and the complex thermodynamics of multi-component formulations. Powder quality attributes such as moisture content and tapped density cannot be measured inline, while the dynamics of the dryer itself (long residence times in the fluidized beds, recycles, frequent changeovers) make the plant hard to model from first principles alone.

Accurate dynamic models of spray dryers would unlock a range of applications, from soft sensing of powder properties to forecasting and model predictive control. Mechanistic models struggle with such agglomeration-related phenomena, while standard machine learning models ignore the structure of the plant, making them brittle and hard to transfer between dryers.

Approach

This ongoing project develops a dynamic modeling framework that embeds the process topology of the spray dryer directly into a machine learning model. The multistage dryer (homogenizer, spray drying chamber, internal and external fluidized beds, and cyclone with fines return) is represented as a directed graph: unit operations become nodes, material streams become directed edges following the physical direction of flow, and sensor measurements are embedded as attributes on the nodes and edges where they physically sit.

A message-passing graph neural network learns spatial relationships across the dryer topology, producing a flowsheet embedding at each timestep. To capture the long residence times induced by the fluidized beds and recycles, these embeddings are passed through a transformer encoder over a look-back window. The resulting model can serve as the backbone for different downstream tasks, including soft sensing of hard-to-measure powder properties, forecasting of process variables, or as a learned dynamic model inside a predictive controller.

Industrial Case Study

The framework is being developed and validated together with Danone, using long-term continuous production data from an industrial multistage spray dryer for infant formula. This setting brings the full complexity of real operation into the modeling task: many product formulations, frequent changeovers between steady-state and transient regimes, sparse and delayed lab measurements, and unmeasured but critical streams such as fines return. Early results suggest that topology-awareness is particularly helpful for capturing process variables governed by mass and energy flows propagating through the dryer, while phenomena dominated by local agglomeration near the nozzle remain harder to capture.

Beyond a Single Dryer

Because the graph neural network does not assume a fixed input structure, the same framework can in principle be transferred between plants with different sensor networks and topologies. A companion line of work on ammonia synthesis loops explores this transfer learning capability on topologically different processes. For spray drying, this opens the door to eventually training a single dynamic model on data pooled from multiple dryers across sites and product lines.

Outlook

Future work focuses on extending these models towards full digital twins of multistage spray dryers, using them for soft sensing, forecasting, and predictive control, transferring trained models across dryers, and closing remaining measurement gaps to better capture agglomeration-driven powder properties.

Theisen, M.F., Dubbelboer, A., Guedes De Brito Neto, H., Meesters, G.M.H., Schweidtmann, A.M. Topology-aware machine learning soft sensors for multistage spray drying using large-scale industrial data. Internation Granulation Workshop, Hamburg (2025)

Theisen, M.F., Meesters, G.M.H., Schweidtmann, A.M. Graph neural networks for soft sensors: Learning from process topology and operational data. Computers and Chemical Engineering, (2026).

Theisen, M.F., Meesters, G.M.H., Schweidtmann, A.M. Transferring graph neural networks for soft sensor modeling using process topologies. ESCAPE-35, Ghent (2025).