Role of High Fidelity Vs. Low Fidelity Experimental Data in Machine Learning Model Performance for Predicting Polymer...

Area of Science:

• Polymer Science
• Materials Science
• Computational Chemistry

Background:

• Polymer-solvent compatibility is crucial for developing new materials and formulations.
• Machine learning (ML) and artificial intelligence (AI) show promise for predicting polymer solubility.
• Model performance hinges on the quality and nature of the training data.

Purpose of the Study:

• To evaluate how experimental data fidelity impacts ML model performance for polymer-solvent compatibility.
• To compare ML classifiers trained on high-fidelity (turbidity-based) versus low-fidelity (visual inspection) datasets.
• To identify key factors influencing the accuracy of ML-based solubility predictions.

Main Methods:

• Two datasets were generated: high-fidelity from turbidity measurements (Crystal16) and low-fidelity from visual inspection.
• Polymers were encoded using one-hot encoding, and solvents using Morgan fingerprints.
• XGBoost classifiers were employed to predict solubility labels (soluble, insoluble, partially soluble).

Main Results:

• Models trained on high-fidelity data demonstrated superior performance, better capturing partially soluble behaviors.
• Quantitative measurements led to clearer class distinctions compared to subjective visual inspection.
• Including temperature as a feature improved prediction accuracy for the low-fidelity dataset.

Conclusions:

• Experimental data rigor is paramount for developing generalizable ML tools in polymer science.
• High-fidelity, quantitative data sources are essential for reliable AI-driven polymer solubility prediction.
• Consideration of experimental parameters like temperature is vital when utilizing literature-derived datasets.