Scope of the In-House CQA Models

Elegent’s in-house modelling framework focuses on the prediction of critical quality attributes (CQAs) relevant to direct compression tabletting. The current models address four key aspects of tablet manufacturability and quality:

• Powder flowability, quantified through tablet weight variability (%RSD)
• Tablet tensile strength, describing mechanical robustness
• Tablet solid fraction, reflecting densification and porosity
• Ejection stress, associated with tablet integrity and tooling wear

Together, these CQAs capture the dominant risks encountered during direct compression, spanning die filling, compaction, decompression, and ejection. Modeling these attributes in an integrated way enables early assessment of formulation and process robustness before extensive experimental work is undertaken.

General Modelling Philosophy

There is no single modeling strategy that is optimal for all CQAs or formulation questions. For this reason, Elegent adopts a case-specific modelling approach, selecting the most appropriate strategy based on:

• The nature of the CQA
• The available data density and structure
• The level of mechanistic understanding available for the underlying phenomena

Where necessary, machine learning methods are used to capture complex, nonlinear relationships between material properties, formulation composition, and process parameters. At the same time, these models are not purely data-driven. Mechanistic knowledge is incorporated wherever possible, either through:

• Physically meaningful input features
• Constraints on model behavior
• Interpretation of outputs in line with known compaction and flow mechanisms

This hybrid philosophy allows the models to balance predictive performance with physical plausibility, ensuring that results remain interpretable and relevant for formulation and process decision-making.

Data Backbone: Experimental Foundation of the Models

The backbone of Elegent’s CQA models is an ever-growing experimental dataset. A substantial part of this dataset originates from a comprehensive database established at Ghent University, for which Elegent is the sole commercial exhibitor. This original database forms a validated and well-characterized foundation for the modeling framework.

Since its inception, the dataset has been continuously expanded by Elegent through additional experimental campaigns. At present, it includes:

• Characterization data for over 80 individual powders, including APIs and excipients
• More than 400 formulations, for which relevant CQAs have been measured

For each formulation, a consistent set of raw material descriptors, blend properties, and process-relevant responses is collected. This growing dataset enables continuous model refinement, extension to new material classes, and improved generalization across formulation space.

Modelling Powder Flow in a Tablet Press

One example illustrating the modelling approach is the prediction of powder flow behavior during die filling, expressed as tablet weight variability.

From Blend Properties to CQA Response

For each formulation, blend-level properties are first predicted based on the properties and proportions of the individual components. These predicted blend descriptors are then linked to experimentally observed tablet weight variation, quantified as the relative standard deviation (RSD) of tablet weight after production on a Styl’One Evo compaction simulator.

Rather than predicting a single continuous RSD value, the flow model is structured as a classification problem. Formulations are classified into two categories:

• Tablet weight RSD ≤ 2%, considered acceptable
• Tablet weight RSD > 2%, considered indicative of insufficient flow

This threshold-based approach reflects common practical acceptance criteria in early formulation development.

Probabilistic Classification

A classifier model is trained on the dataset to distinguish between these two classes based on the predicted blend properties. Importantly, the model does not only provide a binary classification outcome. Instead, it outputs a approximation of a probability of class membership. This outcome expresses the model’s confidence that a given formulation will fall below or above the 2% RSD threshold. As such, it provides:

• A quantitative measure of flow robustness
• A way to compare formulations that lie close to the acceptance boundary
• A practical basis for formulation ranking and design space exploration

Using probabilities rather than hard class labels allows more nuanced decision-making, particularly in early development stages where uncertainty is inherently higher.

Tablet Critical Quality Attributes

In addition to powder flowability, Elegent applies its modelling philosophy to other key CQAs in direct compression, namely tablet tensile strength, tablet solid fraction, and ejection stress. While each of these attributes reflects different physical phenomena during compaction, they share a common methodological backbone rooted in material-level characterization and formulation-aware data integration.

Material-Level Compression Behaviour as Model Input

For tensile strength, solid fraction, and ejection stress, the modelling strategy starts at the individual powder level. APIs and excipients are experimentally characterized with respect to their compression behavior using instrumented compaction experiments. From these experiments, a set of physically meaningful descriptors is derived, including for example:

• Work of compaction and work of compression
• Elastic recovery and decompression behavior
• In-die porosity and densification profiles 
  

These descriptors capture how each powder deforms (plastically or brittly), stores elastic energy, and contributes to bonding during compaction. Rather than treating these measurements as isolated responses, they are decomposed into material-specific scores that describe fundamental compaction mechanisms.

This decomposition step is critical, as it allows the compression behavior of individual powders to be abstracted in a way that can later be recombined when powders are mixed into formulations.

From Individual Powders to Formulations: Neural Network Mixing Strategy

Once material-level compression descriptors are obtained, they are used as inputs to a neural network-based mixing model. This model is responsible for combining the individual powder characteristics into formulation-level predictions.

The neural network thus accounts for nonlinear effects arising from component interactions depending on their properties and presence (mass fractions).

This approach avoids simplistic linear averaging of material properties, which often fails to capture the emergent behavior of powder blends. Instead, the model learns how material-level compaction responses translate into formulation-level behavior in a data-driven yet physically informed manner.

Incorporating Process Conditions

The final stage of the modelling framework links formulation-level descriptors to process settings. The main compaction pressure is the dominant variable influencing tensile strength, solid fraction, and ejection stress, yet additional process parameters are also considered where relevant.

By explicitly including process settings as model inputs, the framework enables prediction of CQAs as continuous functions of operating conditions rather than at a single fixed point. This allows:

• Construction of full tabletability or densification curves
• Identification of pressure regions associated with overcompression or elastic instability
• Evaluation of process sensitivity for a given formulation

Tensile Strength Modelling

For tablet tensile strength, the model predicts the evolution of mechanical strength as a function of compaction pressure. The neural network integrates:

• Powder-level bonding and deformation descriptors
• Formulation composition
• Process conditions

The resulting output is a continuous tensile strength–pressure relationship, enabling identification of plateau behavior and overcompression regions. This provides insight into both achievable tablet strength and mechanical stability limits for a formulation.

Solid Fraction Modelling

The tablet solid fraction model focuses on densification behavior during compaction. Using material-level compressibility descriptors and formulation composition, the model predicts the relationship between applied pressure and resulting solid fraction or porosity.

This enables:

• Quantitative comparison of densification efficiency between formulations
• Evaluation of the link between solid fraction and downstream CQAs such as tensile strength and dissolution
• Identification of pressure ranges where further densification yields diminishing returns

Ejection Stress Modelling

For ejection stress, the modelling framework combines descriptors related to elastic recovery, die-wall friction, and lubricant sensitivity with formulation composition and compaction conditions.

The model predicts ejection stress as a function of compaction pressure and speed, allowing normalization across tablet geometries. By doing so, it supports early identification of:

• Formulations prone to high die-wall friction
• Increased risk of mechanical defects during ejection
• Operating conditions that may accelerate tooling wear

Unified Modelling Framework

Although each CQA model targets a different physical outcome, they share a unified structure:

1. Material-level characterization and decomposition
2. Neural network-based mixing of powder properties into formulation behavior
3. Explicit inclusion of process settings

This consistent framework allows the CQAs to be evaluated in parallel and compared within a common decision-support context. As the underlying dataset continues to expand, these models are further refined, increasing their predictive accuracy and applicability across a wider formulation space.