Accelerating Formulation Design via Machine Learning: Generating a High-throughput Shampoo Formulations Dataset.

Area of Science:

• Formulation Science
• Materials Science
• Chemical Engineering

Background:

• Liquid formulations development is complex and time-consuming due to intricate ingredient interactions.
• Current machine learning (ML) models for formulation design are limited by small datasets and lack of structural information, hindering predictive accuracy.
• Accelerating the design of liquid formulations requires comprehensive datasets and advanced ML approaches.

Purpose of the Study:

• To create a high-dimensional, open experimental dataset for training ML models in rinse-off formulations.
• To expand the ingredient chemical space and design dimensionality for robust ML model development.
• To generate high-fidelity data including phase stability, turbidity, and rheology with sample-specific uncertainty.

Main Methods:

• Curated a dataset of 812 liquid formulations using eighteen diverse ingredients.
• Employed a semi-automated, ML-driven workflow for data generation.
• Measured phase stability, turbidity, and rheological properties for all formulations.
• Incorporated sample-specific uncertainty measurements for advanced model training.

Main Results:

• Developed a dataset with an over 50-fold increase in design space dimensionality compared to previous work.
• Included 294 stable formulations covering the entire design space.
• Generated comprehensive physical property data with associated uncertainty estimates.
• Demonstrated the utility of the dataset for training predictive surrogate models.

Conclusions:

• The presented dataset significantly advances ML-driven liquid formulations design by providing unprecedented scale and detail.
• The inclusion of uncertainty measurements enables the development of more reliable and accurate predictive models.
• This resource accelerates the development cycle for rinse-off formulations, enabling better tuning to target properties.