Predictive modeling of die filling of the pharmaceutical granules using the flexible neural tree

Varun Kumar Ojha, Serena Schiano, Chuan-Yu Wu, Václav Snásel, Ajith Abraham

TL;DR

This paper introduces a Flexible Neural Tree (FNT) model to accurately predict the die filling performance of pharmaceutical granules, a critical step in tablet manufacturing. The FNT model, optimized using genetic programming and differential evolution, outperformed other computational intelligence techniques in predicting the mass of granules deposited into a die, highlighting the potential of advanced machine learning for pharmaceutical process optimization.

What Problem Does This Paper Address?

In pharmaceutical manufacturing, ensuring consistent tablet quality hinges on understanding and controlling various process variables. Die filling, the process of introducing powder or granules into a die before compression, is a crucial stage where variations can lead to inconsistent tablet composition and properties. Identifying the complex relationships between material properties (like granule size and density) and process parameters (like shoe speed) is challenging but essential for developing robust and efficient manufacturing processes. Existing methods often struggle to capture these intricate interactions, necessitating the development of more accurate predictive models.

Methods

The core of this research is the development and application of a Flexible Neural Tree (FNT) model. FNTs are a type of computational intelligence technique that uses genetic programming to create a tree-like structure, similar to feedforward neural networks but with a more adaptable architecture. To enhance the FNT's predictive accuracy, its parameters were further optimized using a differential evolution algorithm. The performance of the FNT model was rigorously evaluated and compared against other established computational intelligence techniques: Multilayer Perceptron (MLP), Gaussian Process Regression (GPR), and Reduced Error Pruning Tree (REP-Tree). The models were trained and validated using two cross-validation methods: 10-fold cross-validation (10FCV) and 5x2-fold cross-validation (5x2FCV).

Datasets

The study utilized experimental data collected from die filling experiments. The dataset comprised:

  • Three grades of microcrystalline cellulose (MCC) powders: Avicel PH-101, Avicel PH 102, and MCC DG.
  • Granules produced by roll-compacting and milling these powders.
  • Six different granule size classes: ranging from 0-90 µm to 1400-2360 µm.
  • Seven different shoe speeds: ranging from 10 mm/s to 400 mm/s.

The input features for the predictive models were:

  1. True density of raw powders
  2. Mean diameter (d50) of raw powders
  3. Granule size (µm)
  4. Shoe speed (mm/s)

The output variable was the mass of granules deposited into the die (g). A total of 389 experimental datasets were generated.

Key Insights and Ideas

  • FNT Outperforms Other Models: The FNT model consistently demonstrated superior predictive performance compared to MLP, GPR, and REP-Tree across both cross-validation methods. This suggests that the FNT's flexible, tree-like structure and its optimization through genetic programming and differential evolution are particularly well-suited for modeling the complex die filling process.
  • Process Variables Dominate: Feature analysis revealed that granule size and shoe speed (process variables) had a more significant impact on the deposited mass than the raw powder properties (true density and d50). This is a crucial insight for process engineers, indicating that controlling these operational parameters is key to managing die filling outcomes.
  • Coarser Granules are Easier to Predict: The study found that the die filling behavior of coarser granules was more predictable than that of finer granules. This is attributed to the higher reproducibility of experimental data for larger granules, suggesting less variability and scattering in their flow behavior.
  • 10FCV vs. 5x2FCV: While both cross-validation methods were effective, 10FCV generally yielded slightly higher correlation coefficients, indicating better model performance on the test sets. However, 5x2FCV was noted for its higher generalization ability due to its distinct training and testing splits.

Key Results

  • The FNT model achieved a high test correlation coefficient of 0.95 in the 10FCV validation, indicating high predictability.
  • In the 5x2FCV validation, FNT models also showed competitive performance, with test correlation coefficients around 0.88.
  • Feature analysis showed that granule size had the highest selection rate (0.86207) and predictability score (0.92563), followed by shoe speed.
  • The models accurately predicted the general trends of deposited mass decreasing with increasing shoe speed and increasing with increasing granule size.

Limitations and Future Work

  • Model Complexity: While FNT models achieved high accuracy, some models (e.g., Model No. 1 in 10FCV) exhibited high complexity (43 nodes), which might be a consideration for practical implementation. The authors also noted that simpler models with feature selection offered a good trade-off between accuracy and complexity.
  • Multi-output Prediction: A fundamental limitation of FNTs, as noted by the authors, is their current inability to handle multi-output prediction problems directly; they are inherently multi-input-single-output models. This would require developing multiple FNT models for each output.
  • Outlier Prediction: While the models generally followed the trend, some outliers in the experimental data were not perfectly predicted, particularly in the 5x2FCV validation (Fig. 6d). This suggests that noise or unusual events during the die filling process can still pose challenges for prediction.

Future work could explore extending FNTs for multi-output scenarios and further investigating the impact of noise and experimental variability on model performance.

Summary updated

BibTeX
@article{ojha2018predictive,
  author = {Varun Kumar Ojha and Serena Schiano and Chuan-Yu Wu and Václav Snásel and Ajith Abraham},
  title = {{Predictive modeling of die filling of the pharmaceutical granules using the flexible neural tree}},
  journal = {Neural Computing and Applications},
  year = {2018},
  doi = {10.1007/s00521-016-2545-8},
}

Loading related content...

Figure