Machine Learning-Driven Cooling Window Design Beyond Hyperbolic Metamaterials.

Area of Science:

• Materials Science
• Optics
• Computational Science

Background:

• Analytical multilayer designs are limited to narrow spectral bands.
• Machine learning (ML) offers potential for optimizing multilayers across multiple bands.
• The performance of ML-optimized multilayers versus analytical designs under identical constraints is largely unproven.

Purpose of the Study:

• To experimentally validate the superiority of ML-driven inverse design for multilayer coatings.
• To develop a cooling-window coating with high average visible transmittance (AVT) and high average near-infrared reflectance (ANR).
• To compare ML-optimized aperiodic designs with periodic hyperbolic metamaterial (HMM) counterparts.

Main Methods:

• Utilized a factorization machine integrated with simulated annealing for ML-driven inverse design.
• Designed and fabricated ZnS/Ag multilayers.
• Benchmarked ML designs against periodic hyperbolic metamaterial (HMM) structures.

Main Results:

• ML-designed coatings achieved superior performance (0.57 AVT, 0.98 ANR) compared to HMMs (0.49 AVT, 0.83 ANR) under a 156 nm thickness constraint.
• An extended ML design (300 nm) reached 0.79 AVT and 0.97 ANR by suppressing Fabry-Perot resonances.
• ML-driven multilayers demonstrated tunable transmitted colors across the visible spectrum, unlike HMMs.

Conclusions:

• ML-driven inverse design is a powerful method for creating high-performance, ultrathin, and color-tunable cooling-window coatings.
• These advanced coatings offer significant potential for urban energy savings.
• The study experimentally confirms the advantage of ML optimization over traditional analytical methods for multilayer optical coatings.