A machine learning study on superlattice electron blocking layer design for AlGaN deep ultraviolet light-emitting diodes using the stacked XGBoost/LightGBM algorithm

Rongyu Lin, Zhiyuan Liu, Peng Han, Ronghui Lin, Yi Lu, Haicheng Cao, Xiao Tang, Chuanju Wang, Vishal Khandelwal, Xiangliang Zhang, Xiaohang LiJournal of Materials Chemistry C 2022 · DOI

Abstract: A stacked XGBoost/LightGBM model was developed to predict and systematically investigate various high-performance SL-EBLs and to suggest a simpler and experimentally realizable low Al-content SL-EBL design.

This research, published in the Journal of Materials Chemistry C, demonstrates how ensemble machine learning can bypass the computationally expensive "trial-and-error" phase of semiconductor device physics to design more efficient deep-ultraviolet (DUV) LEDs.

The Efficiency Bottleneck in DUV LEDs

Deep-ultraviolet LEDs (200–280 nm) are essential for water purification and sterilization, but they suffer from notoriously low efficiency. A primary bottleneck is electron leakage: electrons often "overshoot" the active region and escape into the p-type layers without recombining to produce light.

To mitigate this, engineers use an Electron Blocking Layer (EBL). While a simple "bulk" layer of AlGaN is common, a Superlattice (SL) EBL—consisting of many thin, alternating layers—is much more effective at reflecting electrons while allowing holes to pass through. However, designing an SL-EBL involves a complex design space of layer thicknesses, aluminum compositions, and band offset ratios that interact in non-linear ways.

Schematic of the AlGaN DUV LED structure and the design parameters for the superlattice electron blocking layer (SL-EBL).

Methods: Bridging Physics and Machine Learning

Traditional TCAD (Technology Computer-Aided Design) simulations are too slow to explore this massive design space comprehensively. The authors combined semiconductor physics simulations with a "stacked" machine learning approach to accelerate discovery:

  1. Data Generation: They used APSYS TCAD software to simulate thousands of LED variants, varying the Al-composition ($c_1, c_2$) and thickness ($t_1, t_2$) of the superlattice layers across different band offset ratios.
  2. Stacked Ensemble Learning: Instead of relying on one algorithm, they "stacked" XGBoost and LightGBM. In this architecture, multiple models are trained, and their outputs are weighted and combined to produce a final prediction. This approach is generally more robust and accurate than using a single model.
  3. Optimization: The ML model was used to map the relationship between SL-EBL geometry and Internal Quantum Efficiency (IQE), identifying optimal configurations that would be impossible to find manually.
The machine learning workflow: from TCAD data generation to the stacked XGBoost/LightGBM ensemble model for IQE prediction.

Datasets and Features

The study utilized a synthetic dataset generated via TCAD simulations with a 70/30 training/testing split. The volume included several hundred to a thousand samples for each of the four investigated band offset ratios (0.55, 0.6, 0.65, and 0.7). Input features included the Al-composition and thickness of odd/even layers, while the target output was the IQE.

Key Insights: Navigating the Design Space

  • The "Trap" Risk: If the aluminum content in the superlattice "wells" is too low (specifically, if it matches the composition of the active quantum wells), the SL-EBL stops being a barrier and starts acting like a trap, leading to parasitic recombination that kills efficiency.
  • Compositional Contrast: The optimal design requires a larger difference in Al-composition between the two superlattice layers when the band offset ratio is higher to balance polarization-induced charges against the electron barrier height.
  • The 3nm Rule: Researchers found that an average thickness of 3 to 4 nm for the superlattice layers is the "Goldilocks" zone. Thinner layers allow too much electron tunneling, while thicker layers increase operating voltage and degrade material quality.

Key Results

  • High Predictive Accuracy: The stacked ML model achieved a Root Mean Square Error (RMSE) as low as 0.9% to 1.7%, providing predictions nearly identical to slow TCAD simulations in a fraction of the time.
  • Performance Leap: The proposed optimized SL-EBL (Sample A: 1 nm $Al_{0.7}Ga_{0.3}N$ / 5 nm $Al_{0.58}Ga_{0.42}N$) achieved an IQE and Wall-Plug Efficiency (WPE) roughly 70% higher than conventional bulk EBL designs.
  • Experimental Feasibility: Unlike some theoretical designs that require extreme Al-content, this optimized design uses relatively low Al-content, making it easier to manufacture using standard epitaxial growth techniques.

Limitations and Future Work

While the ML model is highly accurate relative to TCAD simulations, the study remains theoretical. Physical growth and experimental testing are required to validate the real-world performance of the proposed designs. Additionally, while the methodology is transferable, specific "optimal" designs might shift for shorter wavelengths (e.g., 222 nm far-UVC). Future work could integrate p-type doping profiles as an additional variable to further refine hole injection efficiency.

Summary updated

Abstract

A stacked XGBoost/LightGBM model was developed to predict and systematically investigate various high-performance SL-EBLs and to suggest a simpler and experimentally realizable low Al-content SL-EBL design.
BibTeX
@article{lin2022machine,
  author = {Rongyu Lin and Zhiyuan Liu and Peng Han and Ronghui Lin and Yi Lu and Haicheng Cao and Xiao Tang and Chuanju Wang and Vishal Khandelwal and Xiangliang Zhang and Xiaohang Li},
  title = {{A machine learning study on superlattice electron blocking layer design for AlGaN deep ultraviolet light-emitting diodes using the stacked XGBoost/LightGBM algorithm}},
  journal = {Journal of Materials Chemistry C},
  year = {2022},
  doi = {10.1039/d2tc02335k},
}
Figure