The study investigates the application of various neural network-based models for predicting temperature distribution during the freeze-drying process of biopharmaceuticals. This process is particularly important for heat-sensitive biopharmaceutical products, where maintaining the integrity of pharmaceutical compounds during drying is critically necessary.
Using machine learning (ML) techniques, the research explores several neural network architectures, including Single-Layer Perceptron (SLP), Multi-Layer Perceptron (MLP), Fully Connected Neural Network (FCNN), and Deep Neural Network (DNN). These models were optimized using the Fireworks Algorithm (FWA), showcasing their predictive capabilities.
By focusing on temperature distribution prediction, the study addresses significant challenges faced during biopharmaceutical manufacturing. The ability to accurately track temperature fluctuations is integral to ensuring the effectiveness and safety of biopharmaceutical products, which include vaccines, hormones, and proteins.
The results revealed promising performances across all models, with the MLP demonstrating the highest accuracy on both training and test datasets, achieving R² scores of 0.99717 and 0.99713 respectively. This suggests MLP's efficiency in accurately forecasting temperature based on spatial coordinates.
While the SLP also showed commendable performance, with scores of 0.88903, it lagged behind the more complex models like MLP and DNN, which scored 0.99639. The FCNN also performed admirably achieving R² of 0.99158, solidifying the effectiveness of these neural network approaches for the intended task.
The integration of the Fireworks Algorithm for optimization played a pivotal role, allowing for significant enhancements across the various models. This hybrid strategy merging CFD and advanced ML techniques presents new possibilities for accurate modeling of the freeze-drying process.
Given the complex nature of the freeze-drying process, effective temperature control is governed by both mass and heat transfer phenomena, making neural network-based predictions increasingly valuable. The study emphasizes how computational models can simplify this process, enabling stakeholders to make informed decisions during biopharmaceutical production.
Looking forward, the research indicates potential paths for future investigations, particularly the exploration of additional optimization strategies and model architectures to yield even more precise and reliable predictive capabilities.
This lays the groundwork not just for improvements within the biopharmaceutical industry but also highlights broader applications within mathematical modeling and predictive analytics.