RESUMEN
Utilising three artificial intelligence (AI)/machine learning (ML) tools, this study explores the prediction of fill level in inclined linear blenders at steady state by mapping a wide range of bulk powder characteristics to processing parameters. Predicting fill levels enables the calculation of blade passes (strain), known from existing literature to enhance content uniformity. We present and train three AI/ML models, each demonstrating unique predictive capabilities for fill level. These models collectively identify the following rank order of feature importance: RPM, Mixing Blade Region (MB) size, Wall Friction Angle (WFA), and Feed Rate (FR). Random Forest Regression, a machine learning algorithm that constructs a multitude of decision trees at training time and outputs the mode of the classes (classification) or mean prediction (regression) of the individual trees, develops a series of individually useful decision trees. but also allows the extraction of logic and breakpoints within the data. A novel tool which utilises smart optimisation and symbolic regression to model complex systems into simple, closed-form equations, is used to build an accurate reduced-order model. Finally, an Artificial Neural Network (ANN), though less interrogable emerges as the most accurate fill level predictor, with an r2 value of 0.97. Following training on single-component mixtures, the models are tested with a four-component powdered paracetamol formulation, mimicking an existing commercial drug product. The ANN predicts the fill level of this formulation at three RPMs (250, 350 and 450) with a mean absolute error of 1.4%. Ultimately, the modelling tools showcase a framework to better understand the interaction between process and formulation. The result of this allows for a first-time-right approach for formulation development whilst gaining process understanding from fewer experiments. Resulting in the ability to approach risk during product development whilst gaining a greater holistic understanding of the processing environment of the desired formulation.
Asunto(s)
Inteligencia Artificial , Aprendizaje Automático , Redes Neurales de la Computación , Algoritmos , Fenómenos FísicosRESUMEN
The aim of the current study was to characterize the robustness of an integrated continuous direct compression (CDC) line against disturbances from feeding, i.e. impulses of API and short step disturbances. These disturbances mimicked typical variations that can be encountered during long-term manufacture. The study included a primary formulation, with API of standard particle size, which was manufactured at 5 and 10 kg/h production rates, and a modified formulation, with API of large particle size, which was manufactured at 5 kg/h production rate. Overall, the CDC line smoothened all the disturbances, fulfilling the USP uniformity of dosage units (UDU) limit for single tablets. However, runs with the modified formulation failed the pharmacopoeia UDU requirements for the entire run due to high variation between tablets. The primary formulation passed the requirements in all cases. The residence time distribution (RTD) results indicated that the primary formulation allowed better smoothening ability, and an increase in production rate led to poorer smoothening due to shorter RTD. The RTDs revealed that a substantial part of back-mixing took place after the blender. Thus, the tablet press has an important role in smoothening disturbances longer than the mean residence time of the blender, which was very short.