Machine Learning Algorithm for Efficient Design of Separated Buffer Super-Junction IGBT.

Area of Science:

• Semiconductor Device Physics
• Power Electronics
• Materials Science

Background:

• Optimizing the trade-off between turn-off loss (E_off) and on-state voltage (V_on) in Insulated Gate Bipolar Transistors (IGBTs) is crucial for power electronics efficiency.
• Conventional simulation methods face challenges with new design parameters, such as increased buffer doping concentration in separated buffer layer structures.

Purpose of the Study:

• To present an improved IGBT structure with a separated buffer layer to enhance the E_off and V_on trade-off.
• To propose and validate a machine learning (ML) algorithm for efficient parameter optimization and device analysis, overcoming limitations of traditional Technology Computer-Aided Design (TCAD) simulations.

Main Methods:

• Implementation of a novel IGBT structure featuring a separated buffer layer.
• Development and application of a machine learning algorithm for device parameter analysis and optimization.
• Comparative analysis against conventional TCAD simulation tools to assess accuracy and speed.

Main Results:

• The ML algorithm achieved high accuracy, with coefficients of determination (R^2) of 0.995 for V_on and 0.968 for E_off.
• Optimizing the separated buffer concentration led to a 36.2% improvement in E_off, exceeding the ML-predicted 24.7% improvement.
• The ML approach enabled inverse design, yielding four structures meeting target characteristics (E_off = 1.64 μJ, V_on = 1.38 V).

Conclusions:

• The proposed separated buffer structure effectively improves the E_off and V_on trade-off in IGBTs.
• Integrating ML into device analysis offers a strategic advantage for optimizing complex power electronic devices, reducing computational demands.
• The ML-driven approach facilitates precise design, optimization, and reverse engineering of semiconductor devices.