Predicting phosphor particle distribution in white-LED phosphor films using a NAS-optimized physics-constrained...

Area of Science:

• Materials Science
• Optical Engineering
• Artificial Intelligence

Background:

• Phosphor particle distribution critically impacts phosphor-converted white LED (pc-LED) optical performance.
• Current Mie-theory models struggle with realistic non-spherical and agglomerated phosphor particles.
• Experimental screening for optimal LED spectra is time-consuming and costly.

Purpose of the Study:

• To establish a predictive model for phosphor particle distribution based on LED spectral data.
• To overcome limitations of existing models and experimental methods in pc-LED design.
• To accelerate the development cycle for high-performance pc-LEDs.

Main Methods:

• Constructed a spectral dataset using realistic phosphor particles with diverse spectral responses.
• Employed a double-Sigmoid function and film-layered processing for spectral and particle-distribution descriptors.
• Trained a neural network using data augmentation, neural architecture search (NAS), and physics-constrained loss terms.
• Validated the model using Monte Carlo optical simulations and experimental spectral measurements.

Main Results:

• Achieved high R-squared values for inverse-inference of particle number (0.913-0.994) and size (0.831-0.960).
• Demonstrated strong forward spectral consistency (R-squared 0.974-0.983) between simulated and measured spectra.
• Reduced inference time to approximately 1.71 seconds per target, offering orders-of-magnitude acceleration.

Conclusions:

• The developed machine learning approach accurately predicts phosphor particle distribution from spectral data.
• This method significantly accelerates pc-LED design and optimization compared to conventional techniques.
• The physics-constrained AI model offers a cost-effective and efficient alternative for achieving target LED spectra.