(T0930-09-62) An Improved Spray Dried Dispersion Model: Integrating Machine Learning to Predict Optimized Process Parameters

Purpose: When spray drying amorphous solid dispersions (ASD) particle size is an important critical quality attribute to track and maintain to ensure consistent bioavailability and manufacturability of the final spray dried dispersion (SDD). Many solution properties and manufacturing process parameters impact the particle size of the SDD and need to be evaluated when designing a process or scaling up. This can take substantial API, time, and materials/resources. To limit experimental burden a machine learning model to predict particle size independent of API was developed and integrated into an in-house thermodynamic model creating the thermodynamic+ model. The thermodynamic+ model was used to nominate spray drying parameters while predicting the final D50 particle size. Using the nominated process parameters, SDDs were manufactured and actual particle size compared to predicted particle size. All the collected data was then input into a historical database which is integrated into the machine learning feedback loop to iteratively improve future processing parameter predictions.

Methods: A previously developed machine learning model was integrated into an in-house thermodynamic model to create the thermodynamic+ model. Using the improved model, PSD-1 scale spray drying processes were designed to predict both the inlet temperature and the particle size for a given nozzle design. The PSD-1 scale sprays were completed and the particle size distribution was measured using a Mastersizer 3000 (Malvern Instrument Ltd.). The actual volumetric median size (D50) of the SDD were then compared to the predicted value from the thermodynamic and particle size model.

Results: After implementing the thermodynamic+ model, eight spray dried manufactures were completed. The particle size distribution was measured and compared to the predicted size. For the eight manufactures, five sprays predicted within 6 µm of the measured particle size. Sprays that were manufactured with a single solvent had better predictions than sprays from a blended solvent system (13% error vs 44% error respectively). Within the blended solvents, mixtures of DCM and methanol had some of the highest variation from the predicted particle size. When compared to the initial data set used to create and train the machine learning model these observations are not surprising. From the 680 sprays used to develop the model, 45% were of blended solvent systems and only 25 sprays were blended DCM and methanol (3.7% of the data set). The minimal blended solvent and DCM/methanol data integrated into the machine learning leads to a model that has to perform more extrapolation leading to higher variability.

Conclusion: Using the thermodynamic+ model is an ideal way to nominate initial process parameters for amorphous SDDs. By combining thermodynamic modeling and particle size prediction, a set of process parameters can be selected to give a high throughput, optimized process without the large experimental burden. Throughout this research some errors in the model have been assessed. The machine learning model does not predict blended solvent systems as accurately as single solvents, but can be improved. All data from these and future sprays are recorded in an online database which is integrated into the machine learning model. By supplying a greater data set with more experimental values the model will become more accurate and produce better predictions to improve API limited scale up processes.

References: Schmitt, J., Baumann, J. A., & Morgen, M. M. (2022). Predicting Spray Dried Dispersion Particle Size Via Machine Learning Regression Methods. Pharmaceutical Research, 39(12), 3223–3239.