Digital design and operation of continuous crystallization processes via mechanistic modelling

Niall Mitchell, John Mack, Furqan Tahir, Cameron Brown, Muhammad Islam, John Robertson, Eduardo Lopez-Montero, Alastair Florence

Research output: Contribution to conferenceSpeech


Mechanistic models are becoming more commonly applied for Research and Development in the pharmaceutical sector to gain process understanding, enable process design and operation. Traditionally, the output from this activity is a validated mechanistic model, which is capable of quantitatively predicting the behaviour of the various Critical Quality Attributes (CQA) for typical batch or continuous pharmaceutical processes for a wide range of Critical Process Parameters (CPP). However, these tools are almost exclusively employed in an offline manner currently to enable digital design efforts, primarily aimed at assessing process robustness and variability, with very little subsequent online application of the mechanistic model to enable control or soft sensing. Model Predictive Control (MPC) is an established technology in the process industries. It uses a statistical model of the process to capture the dynamic relationships between the inputs (CPPs) and outputs (CQAs) of the process. Using this statistical model, it predicts impact of known disturbances on operation and controls the process through co-ordinated moves on multiple inputs. The MPC exploits all opportunities to reduce variability in the CQAs whilst compensating for measured and unmeasured disturbances. The statisitical process model is built from process response test data at production scale using techniques such as Pseudo-Random Binary Sequence (PRBS) or step-tests. The PRBS test is a relatively non-invasive technique compared to traditional step-tests as it allows the product to remain within specification whilst generating statistically rich information for modelling. Although PRBS testing is suitable for many industries it cannot be used in the Pharmaceutical sector as product generated during testing cannot be utilised for clinical or commercial supply. Consequently, the cost to generate the statistical control model would be significant, presenting a barrier to uptake of the technology. This drawback can be overcome by the integration of mechanistic models, developed using laboratory scale data with MPC system, such as Perceptive Engineering’s PharmaMV platform via a digital design approach. In this work we outline, the application of an advanced process modelling tool, namely gPROMS FormulatedProducts, to describe a number of pharmaceutical crystallization processes. The mechanistic process model and the mechanistic model kinetic parameters were validated using process data gathered from the literature and from lab-based experiments. The lab-based mechanistic model was subsequently used to predict the behaviour of the full scale production scale. The validated mechanistic model was subsequently integrated with PharmaMV to develop and tune the MPC against the mechanistic simulation of the process, by using the mechanistic model as a Digital Twin or Virtual Plant as follows: gPROMS: Build mechanistic model gPROMS: Small scale parameterisation experiments & mechanistic model validation gPROMS + PharmaMV: Validate/check mech model against full scale data gPROMS + PharmaMV: Build MPC using mechanistic model as a digital twin PharmaMV: Transfer MPC to live process and test With this approach, the MPC derived from the mechanistic model was utilized to accurately control the defined CQAs, such as final particle attributes (PSD, yield) for continuous crystallization processes, with reduced material wastage at the production scale
Original languageEnglish
Publication statusPublished - 12 Nov 2019
EventAIChE Annual Meeting 2019 - Orlando, United States
Duration: 9 Nov 201915 Nov 2019


ConferenceAIChE Annual Meeting 2019
Country/TerritoryUnited States


Dive into the research topics of 'Digital design and operation of continuous crystallization processes via mechanistic modelling'. Together they form a unique fingerprint.

Cite this