(218h) Adequacy Testing and Lifecycle Management for a Soft-Sensor Based on State-Estimation Approaches. Case Study: Fluid Bed Granulation

The utilization of deterministic process models for pharmaceutical development is an established practice across the pharmaceutical research and development sector. And despite the current trends in the utilization of artificial intelligence, the application of transport, conservation, and chemistry laws in the derivation of mathematical models (when possible) remains the golden standard. Because models are fit for use, the assumptions and parametrization of these need to be in line with the intended use. And since even the simplest model can easily become overparametrized the practitioner should be cautious of carefully managing physical detail that becomes non-estimable from the measurements and experiments available. Tools like estimability analysis can provide the modeler with useful information about the extent of parametrization that a model can have, based on the available data. Low- and medium-impact deterministic models are then commonly utilized to define operating conditions and ranges for commercial manufacture.

The lifecycle of a deterministic model normally ends at the tech-transfer stage due to the lifecycle requirements for a model-based solution. ICH calls for a model maintenance approach that mostly focuses on ensuring that the accuracy and adequacy of the model is kept with respect to the product lifecycle in manufacturing. These approaches are however insufficient when a model is deployed as a component of a more complex exercise such as that of a state estimator. One of the advantages of using a state estimation approach is the potential to produce accurate estimates for a material attribute using process data only, avoiding the need for an analytical instrument (such as a spectrometer).

A state estimator, such as the Extended Kalman filter (EKF), uses a process model in combination with the available output measurements on an optimization algorithm to update the model state variables in real time such that the produced estimation presents an optimal compromise between the model and measurement uncertainties. When model parameters are estimated simultaneously with state variables, the state estimation optimally adapts the model to resembles the operation accounting for the effect of the unmeasured disturbances. The state estimator then uses measurements from model inputs and measured model outputs to reconstruct all states and parameters including any observable un-measured states (such as an unmeasured material attribute). This calculation does not rely exclusively on the accuracy of a model for a specific set of parameters; but relies more on the model’s structure and flexibility to mimic the physics of the process, irrespective of the specific magnitude of the states within the space where the assumptions hold. As such, model-specific metrics to assess the performance of a state estimator are insufficient and inappropriate to regulate the lifecycle of a state estimator.

In this talk, we will present a set of diagnostics and their use in a proposed life cycle for a soft sensor that is built using state estimation. We illustrate these concepts in the development of a Loss on Drying (LOD) soft sensor using an EKF for a fluid bed granulation.