Control of Powder Based Pharmaceutical Unit Operations Employing Imaging and Machine Learning
Process Analytical Technology (PAT) tools play a vital role in controlling unit operations and transitioning to continuous manufacturing by enabling real-time monitoring of critical quality attributes. The current project focuses on pharmaceutical powder processing unit operations like filtration, drying, blending, and milling. Using imaging and machine learning algorithms, the method ensures accurate particle size measurement, including simultaneous measurement of powder wetness during drying. Ongoing efforts aim to extend this technique to measure blend quality during blending and monitor particle size changes during milling.
Figure: (a) A sketch of our speckle probe. We collect the scattered light from the scattering media—powder in our case—with a monochromatic CCD camera. The Machine Learning-based analysis can extract the quantitative surface information, the PSD, from the speckle statistics. (b) PEACE training loop. The forward operator comes from the physics model. A small “generator” with a physics-picture-inspired structure is trained by a modest amount of the experimental data. The forward operator and the trained generator produce a much larger synthetic dataset. This synthetic dataset trains the DNN “estimator” to learn the mapping from the measured speckle autocorrelation to the particle size distribution (PSD). The ‘generator’ only contains 2.8k parameters, while the estimator has 377k parameters.
