Machine-Learning-Assisted Viscoelastic Characterization of PC/ABS Blends via Multi-Frequency Dynamic Mechanical...

Area of Science:

• Materials Science
• Polymer Physics
• Computational Materials Science

Background:

• Characterizing viscoelastic properties of polymer blends like PC/ABS is crucial for material design.
• Traditional methods like DMA can be time-consuming and require expert interpretation.
• Machine learning offers potential for accelerated and accurate prediction of material behavior.

Purpose of the Study:

• To combine multi-frequency dynamic mechanical analysis (DMA) with machine learning (ML) for characterizing and predicting PC/ABS blend viscoelastic properties.
• To evaluate and compare the performance of various data-driven ML models against a physics-informed NeuralWLF model.
• To establish a quantitative criterion for validation stringency in DMA-ML model evaluation.

Main Methods:

• Multi-frequency DMA temperature sweeps were performed on a PC/ABS blend.
• Data-driven models (RF, XGB, SVR, MLP) and a physics-informed NeuralWLF model were trained and validated.
• A hierarchical validation framework, including temperature-blocked cross-validation and leave-one-feature-out (LOFO), was employed.
• A systematic block size sweep was conducted to investigate validation inflation and establish a gap-to-FWHM ratio criterion.

Main Results:

• DMA yielded a glass transition range of 115.8-123.2 °C and frequency sensitivity of 7.18 °C/decade.
• The physics-informed NeuralWLF model demonstrated superior cross-frequency generalization (R2>0.92) with interpretable Williams-Landel-Ferry (WLF) parameters.
• A physics-data crossover was identified at a gap/FWHM ratio of approximately 2, beyond which NeuralWLF outperformed data-driven models.
• Curriculum learning improved NeuralWLF performance under stringent validation conditions (30 °C validation, R2=0.731).

Conclusions:

• Honest evaluation of DMA-ML models necessitates validation gaps exceeding characteristic feature widths.
• The proposed gap/FWHM ratio serves as a quantitative criterion for assessing validation stringency.
• Physics-informed models like NeuralWLF offer advantages in generalization and interpretability for DMA data, especially beyond the identified physics-data crossover point.