Machine-Learning-Based Dispersion Optimizer for Carbon Nanotubes across Dispersant-Solvent-Process Space.

Area of Science:

• Materials Science
• Nanotechnology
• Computational Chemistry

Background:

• Developing high-quality carbon nanotube (CNT) dispersions is crucial for advanced materials.
• Traditional experimental methods for optimizing CNT dispersion are slow and lack systematicity.
• The combinatorial space of dispersants, solvents, and processing conditions is vast and complex.

Purpose of the Study:

• To develop a machine-learning (ML) model for predicting CNT dispersibility and structural integrity.
• To create an efficient CNT dispersion optimizer that considers multiple formulation and processing variables.
• To reduce the empirical trial-and-error in designing stable CNT dispersions.

Main Methods:

• Collected a dataset of 666 dispersions with systematic variations in dispersants, solvents, and processing.
• Utilized molecular descriptors, experimental variables, and similarity metrics as ML model inputs.
• Employed an eXtreme Gradient Boosting (XGBoost) model to predict dispersion quality (Dscore) and structural integrity (IG/ID).

Main Results:

• The XGBoost model achieved R² = 0.57 (MAE = 0.08) for Dscore and R² = 0.73 (MAE = 9.84) for IG/ID.
• Model performance demonstrated mean absolute errors below 10% of the target ranges, suitable for screening.
• Identified key factors: solvent-dispersant compatibility influences Dscore, while process intensity dominates IG/ID.

Conclusions:

• The ML-based optimizer efficiently prescreens formulation and processing conditions for CNT dispersions.
• This framework significantly reduces experimental effort and accelerates the design of nanomaterial-based formulations.
• The approach is extensible to other nanomaterial dispersion systems and downstream manufacturing processes.