Deep Learning methods for solid loading prediction from in-line Chord Length Distributions (CLDs)

One of the greatest challenges of the pharmaceutical industry is its ongoing transition towards continuous manufacturing, motivated by the higher economic profit, and lower environmental impact over batch processing [1, 2]. Moreover, the enhanced agility of continuous processes enables them to potentially meet the growing global drug demand. The development of reliable in-line process analytical technologies (PAT) are paramount for the implementation of automatic control systems and quality control, two essential components of a continuous plant. FBRM is one of the main in-line characterisation techniques used within industry due to its versatility: it can also be applied off-line, which can help smoothen the transition between both frameworks, and it can be applied to systems up to 40% v/v [3]. This technique yields the Chord Length Distribution of the system (CLD) which constitutes an indicative of its concentration.
In this context, this research aims to develop a Convolutional Neural Network (CNN) regression model capable of inferring solid loading from FBRM measurements. The method has been tested on three different material samples: 1) standard polystyrene spheres, consisting of 3492 CLD samples in a range of 200-400 µm and 1-20%wt, 2) a mixture of polystyrene spheres and ellipsoids, with 3700 CLD in the range 0-500 µm and 1-10%wt, and 3) lactose particles with concentrations from 2.5 to 20%wt and particle sizes from 90 to 500 µm.
The FBRM data is processed with a 1D CNN, of which a diagram is provided in Figure 1, formed by several ResNet [4] convolutional blocks followed by fully connected layers. The model is trained for 500 epochs using a cyclical learning rate approach [5], and several experiments are performed to determine the best combination of hyperparameters for it. More specifically, the number of convolutional and linear layers will be tested, as well as the pooling and kernel size and the value of the maximum learning rate.
Figure 2 shows some preliminary results obtained with the polystyrene spheres dataset. These suggest that R2 values greater than 0.95 and root-mean-square errors of less than 0.3%wt can be achieved for samples the model has not seen during training. These metrics show the potential that CNNs uphold for the development of data analysis tools for in-line PAT.