New paper: Graph neural networks for soft sensors that learn from process topology

Our paper "Graph neural networks for soft sensors: Learning from process topology and operational data" by Maximilian F. Theisen, Gabrie M. H. Meesters and Artur M. Schweidtmann was published in Computers & Chemical Engineering.

Soft sensors estimate process variables that are difficult or impossible to measure directly, such as product concentrations, from available sensor data. Standard machine learning soft sensors learn from historical data only and ignore basic process information such as the plant topology. This can lead to models that capture correlations rather than causal relations, that degrade in unseen operating scenarios, and that need large amounts of data.

The paper proposes process topology-aware graph neural networks. Process data are represented as a directed graph of the plant, in which unit operations are nodes, streams are edges and sensor measurements are attributes. Compared with standard black-box models, the approach gave more robust models, required less data and produced more intuitive data representations. Embedding process information directly into the data is a step towards more reliable digital twins.

We apply this approach in an ongoing project on industrial multistage spray drying together with Danone.

Paper: M. F. Theisen, G. M. H. Meesters, A. M. Schweidtmann (2026). Graph neural networks for soft sensors: Learning from process topology and operational data. Computers & Chemical Engineering, 206, 109532. https://doi.org/10.1016/j.compchemeng.2025.109532