Cross-material physics-informed machine learning framework for optimizing nanofiller loading in epoxy nanocomposites...

Area of Science:

• Materials Science
• Electrical Engineering
• Computational Science

Background:

• Epoxy-based nanocomposites are crucial for high-voltage insulation due to their excellent properties.
• Optimizing nanofiller loading for maximum dielectric breakdown strength (BDS) is challenging using traditional methods.
• Current methods are costly, time-consuming, and lack generalizability.

Purpose of the Study:

• To develop a physics-informed machine learning framework for optimizing epoxy nanocomposite insulation.
• To integrate diverse nanofillers (Zn/Al-LDH, Mg/Al-LDH, γ-Al2O3, α-Al2O3) into a unified predictive model.
• To enable scalable and cost-effective optimization of dielectric properties for GIS/GIL spacers.

Main Methods:

• A cross-material machine learning framework was developed, incorporating four distinct nanofillers.
• Dataset augmentation techniques (PCHIP interpolation, Gaussian noise) were employed for enhanced model training.
• Composite relative permittivity was estimated and integrated as a physics-based input feature using the Maxwell-Garnett model.

Main Results:

• The α-Al2O3 nanofiller yielded the highest BDS (46.8 kV/mm at 5 wt%), a 56% improvement over neat epoxy.
• Material-specific models achieved high accuracy (R² up to 0.959) with low experimental validation errors (<5.2%).
• A transferable cross-material model demonstrated strong performance (R² = 0.919) with prediction errors below 6%.

Conclusions:

• The proposed physics-informed machine learning framework effectively optimizes epoxy nanocomposites for high-voltage insulation.
• The framework offers a scalable, cost-effective, and experimentally validated approach for material design.
• This AI-driven strategy accelerates the development of advanced insulation materials for electrical applications.