Introduction
Spray freeze drying forms a frozen pellet from liquid under cryogenic conditions as droplets are formed and pass through a tower. Subsequently, the frozen pellets as a bulk bed of particles are dynamically lyophilized in a rotating drum. Energy application from an infrared source, the controlled temperature of the drum surface and vacuum conditions allow the drying of the pellets under freezing conditions (Figure 1). The process has distinct advantages in the product format, faster processing time, and lower energy consumption compared to a standard lyophilization process. (1, 2).
The creation of solid pellets in bulk (Figure 2), allows for room temperature stability and the opportunity to ship bulk drug substance without cold chain constraints. The pellet nature of the product is good for flexible filling of different doses and combination of different products. While most often considered for pharmaceuticals and diagnostics, the process has been used in a wide range of products and could even be considered for nutritional ingredients or foods where a high level of the product characteristics must be maintained during processing and volumes are limited (3, 4).
Disruptive technologies such as spray freeze-drying are naturally first evaluated in terms of their feasibility with specific products. The growing pressure to shorten development timelines means that applying process understanding is expected to limit experimentation and accelerate timelines. The drive to understand and accelerate continues throughout development into commercialization such that the transferability, scalability, and robustness must be considered earlier in the development and with predictive tools. Data science including mining of existing data sets, analysis of data for insight, and modeling approaches can add value by minimizing the risk for process failures by increasing the process understanding and insight.
Recent advances in computing capabilities and artificial intelligence have allowed the growth of cloud-based platforms to analyze processing and formulation data (5). These platforms often give advanced graphic generation to visualize the data and several “what if’ scenarios. Data can be either from the data set or projections.
The following information presents analysis of a dataset generated from freeze-drying experiments. A cloud-based platform (Sunthetics) was employed to visualize the data, build predictive models, and suggest optimal operating conditions for the innovative spray freeze drying process. Due to proprietary constraint the full data set cannot be presented, however details regarding the type of information processed and the use of the analysis and modeling platform are disclosed.
Materials and Methods
A database of the dynamic freeze drying process data from spray freeze drying experiments conducted at Meridion Technologies was assembled. Two sizes of process equipment, the LyoMotion LAB unit (LMLAB)and a pilot scale LyoMotion 30 (LM30) unit were considered. The types of materials ranged from low solids content (~5%) to higher solids content (in excess of 15%) materials ranging from small molecules to large molecules and typical excipients (e.g., sucrose and mannitol). Over 80 processing experiments were included in the database. Due to proprietary constraints further detail regarding the database contents cannot be presented.
The data selected for analysis was a subset of that available. The equipment size, charge of frozen pellets, solid fraction and solid content were chosen as input conditions (only two of these are independent, the third for convenience in looking at the data). Therefore, the analysis was agnostic to the materials type. The data was smoothed to give a single parameter for the vacuum pressure, the chamber wall heat transfer fluid (silicone oil), the infrared power input, and rotation speed. Outputs entered into the set included primary drying time, yield, and moisture content.
The data was further cleansed (removal of blanks, reformatting, and renaming of variables) to prepare it for analysis. The Sunthetics platform (www.sunthetics.io) was then used in three stages: first, to perform exploratory data analysis and generate graphical representations such as histograms and correlation matrices; second, to build predictive machine learning models that related process inputs to outputs; and third, to apply optimization tools that identified optimal operating regions and suggested next experiments. Based on these results, the platform’s design-of-experiments module was employed to generate experimental plans for further exploration of the design space.
Results and Discussion
The assembly and cleansing of the data were typical to a data set coming from several sources in that it required a significant effort to review, repair, reformat, and align on the data to be analyzed. Generation of the model went smoothly with upload, limit setting and model generation available within a few hours. Additional time was spent reviewing the model as the amount of information generated was considerable and to fully understand the significance involved some thought and discussion.
As a first assessment of the data set the histograms of the frequency for different variables was reviewed. Some examples are given in Figure 3 (once again for proprietary reasons the full data cannot be shared). This gives an assessment of how the model may be influenced by a significant representation in the data set at one value or another. From the histograms presented it would be reasonably expected that predictions for the laboratory sized equipment (LMLAB) might be more accurate than the pilot unit (LM30). Similarly, if a product of interest was expected to have a long primary drying time the data set would be lean in this area and prediction might have more variance.
The data correlation matrix for the entire data set, with respect to the primary drying output and with respect to the yield are shown in figure 4. Interactions for primary drying are stronger than for yield such that optimizing for primary drying time can be expected to be more complex than optimizing for yield.
Other representations addressed the importance of specific variables in two other graphical formats when optimizing for a particular output parameter. The data was interesting in that the relative importance was not always what the team expected and that the importance had significant differences in the rank order of importance depending on the output. A three dimensional graph showing the hypothetical system response in the primary drying time output for silicone oil temperature, pressure, and infra red power is shown in Figure 5. The selection of the other variables occurs by selection with convenient slide bars on a graphical interface in the software. The graph demonstrated an unexpected result in that some areas of the operating regime are insensitive to variable input. If the desired result falls within this area, then the process can be shown to be quite robust to process variation.
An optimization tool within the Sunthetics platform was used to identify optimal operating criteria when balancing two outputs simultaneously. For example, a trade-off analysis between primary drying time and yield was generated, as shown in Figure 6. In addition, model diagnostics were reviewed, including error estimates for the predictions, which are illustrated in Figure 7, to help assess model quality and uncertainty.
The graph showed that only part of the training set overlapped the area of interest according to the model predictions. These training data were run with other outputs in mind or were non-optimal estimates showing the potential savings by modeling throughout the experimental process.
Given the difference in the predicted area of interest to optimize the primary drying and yield outputs and the training data set the platform function to generate a design of experiments to explore the region further was tested. Designs with as few as 6 runs and more if needed were suggested based on the number of variables and outputs desired for study. An 8 run design is under consideration by the processing study team.
Conclusions
Spray freeze drying is an innovative processing technique with significant potential across industries, whose full range of applications and processing efficiencies are still being explored. Its development requires systematic exploration of complex design spaces. In this study, we demonstrated how a cloud-based AI platform (Sunthetics) can accelerate that exploration by combining exploratory data analysis, predictive modeling, and Bayesian optimization tools in one place. The study demonstrates how new AI-ML visualization, modeling and optimization tools are effective means for scientists to glean knowledge from existing data sets and more efficiently plan experiments and select optimal conditions.
Figure 1. Lab/Pilot Scale Spray Tower and Rotary Dryer for Spray Freeze Drying (courtesy Meridion Technologies).
Figure 2. SEM of a single lyophilized pellet and photo of a group of lyophilized pellets (courtesy Meridion Technologies)
Figure 3. Histograms showing the distribution of data in the selected dataset for model building. Histograms for Equipment Size and Primary Drying Time shown as examples.
Figure 4 (A,B,C). The data correlation matrix for the entire data set and related to two outputs: primary drying and yield.
Figure 5. System response with respect to primary drying time as a function of silicone oil temperature, pressure, and infra red power input with selection of the other input variables to fixed conditions selected via graphical interface.
Figure 6. Desired operating space when optimizing for primary drying time and yield.
Figure 7 (A,B). Estimation of model error for primary drying and yield outputs.
References and notes
- Luy, B. and Stamato, H. (2020). Spray Freeze Drying. In Drying Technologies for Biotechnology and Pharmaceutical Applications (eds S. Ohtake, K.-i. Izutsu and D. Lechuga-Ballesteros). https://doi.org/10.1002/9783527802104.ch8
- Luy, B., Plitzko, M., Stamato, H. (2023). Design and Process Considerations in Spray Freeze Drying. In: Jameel, F. (eds) Principles and Practices of Lyophilization in Product Development and Manufacturing . AAPS Advances in the Pharmaceutical Sciences Series, vol 59. Springer, Cham. https://doi.org/10.1007/978-3-031-12634-5_14
- A. Steegmans, M. Plitzko, B. Luy, S. Lebeer, F. Kiekens, Spray freeze drying as a novel drying process for the formulation of probiotic powders containing Lacticaseibacillus rhamnosus GG, European Journal of Pharmaceutics and Biopharmaceutics, Volume 212, 2025, 114748, ISSN 0939-6411, https://doi.org/10.1016/j.ejpb.2025.114748.
- S. Padma Ishwarya, C. Anandharamakrishnan, Andrew G.F. Stapley, Spray-freeze-drying: A novel process for the drying of foods and bioproducts, Trends in Food Science & Technology, Volume 41, Issue 2, 2015, Pages 161-181, ISSN 0924-2244, https://doi.org/10.1007/s10068-023-01409-8
- H. Stamato, D. Blanco, Machine Learning and Artificial Intelligence for Formulation Optimization 2024 Annual Meeting paper 184u, American Institute of Chemical Engineers, November 2024. https://proceedings.aiche.org/conferences/aiche-annual-meeting/2024/proceeding/paper/184u-machine-learning-and-artificial-intelligence-formulation-optimization accessed August 18, 2025.
ABOUT THE AUTHOR
Howard J. Stamato has over 30 years of experience in a broad range of skills. He has worked with projects from most major pharmaceutical companies, as well as generic and nutritional products. His work has been instrumental in delivering many life-changing medications with annual sales in the billions. Working for two major chemical companies, in household and personal products, with a multinational equipment manufacturer and clients from other industries rounds out his knowledge. Of interest during that career were: designing equipment, facilities, processes, products, and business workflows. He has been a key contributor advancing information technology systems, simulation/modeling, and has pioneered in fields such as knowledge management, data science, and visualization. He has also worked in regulatory affairs, as a consultant, served as an expert witness, and contributed on several advisory boards. Howard has participated in over 60 peer reviewed presentations, posters, and articles as well as patents and book chapters.
