Advancing Direct Tablet Compression with AI

See the poster, which was presented at the 8th Solid Oral Dosage Form2025 Conference:

Introduction

Pharma 4.0 integrates automation, digitalization, and AI to enhance manufacturing efficiency. Direct Tablet Compression (DTC) is complex due to material, formulation, and process parameters. AI-driven approaches support optimization, sustainability, and root-cause analysis.

Objectives

• Predict tablet/batch quality attributes.
• Accept or reject tablets/batches automatically.
• Provide explainability for decisions.
• Develop a multi-task model (Neural Network + LLM).
• Validate using Harvard Dataverse dataset.

Dataset

This research uses 1,982 Fast-Disintegrating Tablet (FDT) formulations from the Harvard Dataverse (2010–2020) [5]. Each entry contains 79 molecular, structural, and compositional features, plus four target quality attributes: disintegration time, friability, hardness, and water absorption ratio.

Preprocessing

• Outlier Labeling: Acceptance/rejection flags based on Pharmacopeia thresholds.
• Data Split: 80/10/10 (train/validation/test).
• Feature Scaling: Standardized features.
• Augmentation: 5× synthetic samples with Gaussian noise (scaled by feature importance).
• Target Noise: Added controlled noise to regression targets for robustness.
• Distribution Check: Verified augmented data preserved original statistics.

Method

• Unified AI Framework for Direct Tablet Compression combining: statistical analysis, machine learning, neural networks, and generative AI (T5).
• Enhancements: enables real-time monitoring, quality prediction, and automated reporting.
• Architecture (Fig. 1):
  • Attribute Regressor: shared backbone for regression (4 attributes) + classification (accept/reject)
    prediction.
  • T5-small: generates cause-based textual explanations from metrics + classification outputs (Why
    accept/Reject batches).

• Modeling Details (Fig. 2): feedforward layers, batch normalization, dropout, and multi-head attention. Trained with early stopping and fine-tuning.
• Text Generation with T5: fine-tuned via LoRA, optimized with cross-entropy loss.
• Key Advantage: multitask learning improves generalization and strengthens prediction–explanation consistency.

Results

Table 1. Model Performance: Regression, Classification, and Generation Tasks

DT: Disintegration Time, WAR: Water Absorption Ratio, TVR: Tuned Voting Regressor, TRF: Tuned Random Forest, TKNNR: Tuned KNN Regressor, DNN: Dense Neural Network, ETR: Extra Tree Regressor, TDA: Tabular Data Augmentation, MHA-DNN: Multi-Head Attention DNN, T5-small: Text-to-Text Transfer Transformer.

• Performance: Avg. R2 = 0:918, RMSE = 0.32 (with augmentation)
• Advantage: Outperforms Momeni et al. and Gupta et al. across all targets
• Unified Model: Captures attribute interdependencies ! more accurate than separate models
• Explainability: via Fine-tuned T5-small (16 epochs) generates interpretable justifications
• Limitations: Missing tablet dimensions & powder properties
• Impact: Demonstrates AI’s potential to enhance prediction, classification, and explainability in tablet manufacturing

See the full technical poster on Advancing Direct Tablet Compression with AI here

(click the picture to download the poster)

Source: Yazid BOUNAB, Osmo Antikainen, Mia Sivén, Anne Juppo, poster: Advancing Direct Tablet Compression with AI, University of Helsinki