Machine Learning to Predict Junction Temperature Based on Optical Characteristics in Solid-State Lighting Devices: A...

Area of Science:

• Solid-state lighting
• Optoelectronics
• Materials science

Background:

• Junction temperature control is crucial for reliable solid-state lighting.
• Direct measurement of junction temperature in LEDs is not feasible.
• Existing methods rely on temperature-sensitive optical parameters and spectral analysis.

Purpose of the Study:

• To demonstrate that color characteristics and power level data are sufficient for predicting white LED junction temperature.
• To develop and evaluate machine learning models for junction temperature estimation.
• To provide a novel, reliable prediction tool for scenarios with measurement limitations.

Main Methods:

• Utilized a database of manufacturer datasheets for white LEDs.
• Developed four machine learning models: k-Nearest Neighbor (KNN), Radius Near Neighbors (RNN), Random Forest (RF), and Extreme Gradient Booster (XGB).
• Trained and tested models using dynamic opto-thermal measurements, correlating spectral power distribution (SPD) shape (Correlated Color Temperature - CCT) and intensity (luminous flux) with junction temperature.

Main Results:

• Machine learning algorithms proved effective in predicting junction temperatures.
• The developed models accurately estimated junction temperatures based on color and power data.
• The approach is particularly valuable for in-situ measurements and challenging environments.

Conclusions:

• Machine learning offers a reliable and novel approach for estimating white LED junction temperature.
• This method overcomes limitations of direct measurement, enabling accurate predictions in diverse applications.
• The findings support the use of ML for junction temperature estimation in wafer-level probing and with phosphor-coated chips.