Machine Learning-Based Analysis and Optimization of Temperature Distribution in Pharmaceutical Freeze Drying: Many temperature-sensitive products involved in the biopharmaceutical industry exhibit limited stability if kept in aqueous solution for a long time. This study investigates the application of various neural network-based models for predicting temperature distribution in the freeze-drying process of biopharmaceuticals.
Machine Learning-Based Analysis and Optimization of Temperature Distribution in Pharmaceutical Freeze Drying
Biopharmaceutical products have gained more attention as the major therapeutics for various disease treatments. Some examples of biopharmaceutical products have gained more attention as the major therapeutics for various disease treatments. Since these materials are sensitive to heat, normal drying by heating is not applicable for biopharmaceuticals. The method of freeze drying is preferred which does not apply heat for removal of moisture and dying is preferred which does not apply heat for removal of moisture and dying the products.
Design and optimization of the freeze-dying process are challenging as both mass and heat transfer phenomena must be taken into account for analysis of the process. Indeed, the temperature must be taken into account for the analysis of the process, During this process Temperature and concentration change over time, They must be controlled to meet the required specifications.
The field of machine learning has gained considerable attention in recent decades due to its broad range of applications in various domains.
Pharmaceutical Freeze Drying
The process of freeze-drying can be well understood and optimized via the development of predictive computational models. It is based on mass and heat transfer can be developed to track the changes in moisture and concentration over time.
Lyophilization is a time-consuming and energy-demanding process that consists of three main steps: Freezing, primary drying and secondary drying. In the first step, the product is frozen at a very low temperature (Around-50 °C) and most of the water solvent is converted to ice. in the second step, ice is removed from the frozen product by direct sublimation at low temperature (around- 20 to -30 °C) and low pressure (Around 50 -100 mTorr) In the third step, the unfrozen water that did not crystallize during freezing and that is bound to the product is desorbed at higher temperature (Around 20 to 30 °C)
Dataset and process
The process of freezing drying is simulated with CFD and machine learning models. The goal is to obtain the temperature distribution inside the dryer at different locations. For the CFD simulation, COMSOL Multiphysics 3.5 software was used which operates based on finite element scheme. Molecular diffusion and conduction were considered for mass transfer and heat transfer respectively.
Computation and Modeling
The datasets for machine learning are obtained by numerical simulation of heat and mass transfer on the simple 3D domain, and the data was extracted for machine learning models. It is important to prepare the data adequately through pre-processing steps.
- Outlier Detection Using the analysis and interpretation of data. By measuring the impact of each data point on the regression coefficients, Cool`s distance helps identify observations with substantial leverage on the model.
- Normalization using Z-score Normalization method: The process of normalization is important to maintain thereby preventing any individual feature from exerting excessive influence on the analysis as a result of its larger magnitude.
Fireworks Algorithm (FWA)
It stands as an optimization technique inspired by the bursts of fireworks to generate sparks, There are three basic operations – Exploration, variation and selection make up the method.
The optimization process utilizing the FWA is centred around the fitness function defined as the mean R2 score obtained from a 5fold cross-validation. This metric ensures a robust evaluation of the model performance, minimising overfitting and capturing generalization across different data partitions.
Result and Discussion
the performance of four neural network-based models. namely, SLP, MLP FCNN, and DNN were evaluated for output temperature forecasting using a dataset containing over 55,000 data points.
The optimized MLP model emerged as the top-performing neural network for spatial temperature forecasting, achieving exceptional predictive accuracy across multiple evaluation metrics. Specifically, the MLP achieved a mean R score of 0.99717 on the training dataset and 0.99713 on the test dataset, demonstrating its ability to explain over 99% of the variance in temperature based on coordinates.
Conclusion
This study demonstrates the effectiveness of neural network-based models including the SLP, MLP, FCNN, and DNN in predicting temperature based on spatial coordinates for the simulation of the freeze-drying process for biopharmaceuticals.