(92h) Soft Sensors in Pharma Applications: Observer Design and Practical Case Studies

The pharmaceutical industry is gradually moving from traditional operating modes to more advanced approaches. This transition is driven by the need to meet changing market demands and comply with strict regulations - such as ensuring the consistent delivery of high-quality products in adequate quantities - and improve production flexibility while reducing costs. A prominent example of this shift is the increased adoption of continuous manufacturing over batch-based processes. In this context, measuring critical quality attributes (CQAs) is essential, as it enables: [a] advanced process control to dynamically optimize critical process parameters and maintain CQAs at target values, [b] real-time quality control to enable diversion of non-conforming material, and [c] end-product quality assessment facilitating real-time release testing. Process Analytical Technology (PAT) tools - typically referring to hardware systems physically integrated into production lines - are well-established in the pharmaceutical industry. In recent years, however, soft sensors have gained increasing attention as virtual, model-based alternatives to conventional hardware-based approaches [1]. This presentation outlines their fundamental principles, design workflows, and application results, illustrated through practical case studies. The design methodology is initially demonstrated through a soft sensor developed to predict loss-on-drying (LOD) in a ConsiGma™-25 fluid-bed dryer (FBD). The workflow is then extended to two additional case studies: one for predicting product concentration in a fed-batch bioreactor, and another for predicting hold-up in a continuous powder blender. The talk will cover the following aspects:

• Reasoning for selecting soft sensors over conventional hardware-based PAT tools, such as the lack of suitable measurement techniques, physical limitations, or high associated costs. Specifically, in the FBD study, the quantity of interest is the LOD across six dryer cells. However, physical integration of an NIR probe is feasible in only two of these cells. Therefore, hardware-based PAT applications would require significant equipment restructuring and additional costs (e.g., purchasing additional NIR probes).
• Preliminary requirements: Since soft sensors rely on process data and models to estimate unmeasured variables, the focus will be on selecting appropriate models – such as computationally lightweight models capable of running in parallel with the process - on acquiring necessary data, including model inputs and outputs in real-time, and checking the observability criteria. In the FBD case study, the model is based on mass and energy balance differential equations, with states representing the mass of dry and wet granules, air moisture content, and the internal energy of air and granules [2].
• Observer design: Two strategies will be compared: a trivial observer, which is an open-loop implementation of the process model commonly used in pharmaceutical applications but often questioned for its sensitivity to model uncertainties and deviations in initial conditions; and an advanced observer, which uses measured model outputs as feedback to improve prediction accuracy. For example, in the FBD study, granule temperature measurements are used to refine the prediction of states required for LOD calculation.
• Performance tests: Finally, practical performance challenges - particularly in the presence of material variability and process disturbances - will be discussed. In the FBD study, these challenges are demonstrated using real process data, which includes intentional variations in the liquid-to-solid ratio during the wet granulation stage. The resulting impact on both the underlying process model and the accuracy of the soft sensor’s LOD prediction will be examined.

_________________________________________________________________________________________________________________________________________________________

[1] Garcia-Munoz et al., Adequacy Testing and Lifecycle Management for a Soft-Sensor Based on State Estimation Approaches. Case Study: Fluid Bed Granulation, AIChE annual meeting, 2024

[2] Rehrl et al., End-Point Prediction of Granule Moisture in a ConsiGma^TM-25 Segmented Fluid Bed Dryer, Pharmaceutics, Volume 12, 2020, https://doi.org/10.3390/pharmaceutics12050452.