Machine Learning-Based Figure of Merit Model of SIPOS Modulated Drift Region for U-MOSFET

Zhen Cao, Qi Sun, Chuanfeng Ma, Biao Hou, Licheng JiaoMicromachines 2024 · DOI

TL;DR

This paper introduces a novel machine learning-based model for optimizing Superjunction (SJ) U-MOSFETs by using semi-insulating poly-crystalline silicon (SIPOS) pillars to modulate the drift region. This approach allows for a better trade-off between breakdown voltage (BV) and specific ON-resistance (RON,sp), pushing beyond the conventional SJ silicon limit.

What Problem Does This Paper Address?

Power MOSFETs are critical for energy conversion, and Superjunction (SJ) technology has been key to achieving high breakdown voltage (BV) and low specific ON-resistance (RON,sp). However, conventional SJ devices face limitations in further optimizing this trade-off, particularly due to electric field distribution issues beneath the trench gate and charge balance challenges. Existing methods like using high-K dielectrics or trench structures have their own complexities. This paper aims to overcome these limitations by proposing a new device structure and a sophisticated modeling approach to achieve superior performance.

Methods

The core of this work lies in a hybrid approach combining physics-based analytical modeling with machine learning.

  1. Analytical Model Development: The authors first derive a physics-informed analytical model for the Superjunction U-MOSFET with SIPOS pillars (SSJ-UMOS). This model accounts for:
  2. Electric Field Modulation: How the SIPOS pillars influence the electric field distribution in the drift region.
  3. Charge Coupling: The interaction between the SJ structure and the SIPOS/oxide/silicon (MIS) structure.
  4. Majority Carrier Accumulation: The effect of a positive gate bias on the MIS structure.
  5. Poisson's Equation: Solved to determine the electrostatic potential and electric field distribution.
  6. Machine Learning for Optimization: Gaussian Process Regression (GPR) is employed as the machine learning technique.
  7. Figure of Merit (FOM): The key performance metric, defined as FOM = BV²/RON,sp, is targeted for prediction.
  8. Hyperparameter Optimization: GPR is used to optimize crucial device parameters (represented by coefficients α, β, λ) that govern the SIPOS pillar's influence.
  9. Dataset Generation: TCAD simulations are used to generate a dataset of device characteristics for training the GPR model.

This combined approach aims to provide an accurate predictive model that can guide the design of optimal device structures.

Datasets

The paper does not introduce a new benchmark dataset. Instead, it relies on:

  • TCAD Simulation Data: A dataset of 1000 samples generated through Sentaurus TCAD simulations, covering various combinations of structural parameters (LD, ND, tox, WN). This data is used to train and validate the GPR model.
  • Experimental Data: The authors calibrate their TCAD simulations and analytical model against experimental breakdown characteristic (ID-VD) data from fabricated SJ-VDMOS devices [21] for validation.

Key Insights and Ideas

  • SIPOS Pillars as E-field Modulators: The SIPOS pillars beneath the trench gate are not just passive elements; they actively modulate the electric field, enabling a more favorable trade-off between BV and RON,sp. This is a key structural innovation.
  • Breaking the SJ Silicon Limit: The proposed SSJ-UMOS structure, through the combined effects of SIPOS modulation, charge coupling, and carrier accumulation, demonstrably surpasses the theoretical SJ silicon limit for FOM.
  • Hybrid Modeling Synergy: The combination of a physics-based analytical model with GPR for hyperparameter optimization is a powerful strategy. The analytical model provides physical intuition and constraints, while GPR efficiently explores the parameter space for optimal performance.
  • Hyperparameter Sensitivity: The study highlights that certain parameters (like α) have a significant impact on fitting results, guiding future optimization efforts.

Key Results

  • Improved Breakdown Voltage: The SSJ-UMOS structure achieves a higher breakdown voltage (e.g., 725 V) compared to conventional SJ-UMOS (e.g., 607 V) for similar device dimensions.
  • Reduced ON-Resistance: The SIPOS-assisted depletion of N-pillars in SSJ-UMOS leads to a reduction in RON,sp.
  • Accurate Prediction Model: The GPR model demonstrates high predictive accuracy, with a Mean Absolute Percentage Error (MAPE) of only 4.5%.
  • Optimal Parameter Identification: The GPR model successfully identifies optimal hyperparameters (α ≈ 0.85, β ≈ 0.53, λ ≈ 0.78) and provides guidance on optimal structural parameters like oxide thickness (tox) and N-pillar width (WN) for different BV targets. For instance, a 1000 V device requires a tox of 0.05 µm and a mesa width of 1.0 µm.
  • Dynamic Performance: While turn-on speed is comparable, the SSJ-UMOS exhibits slightly slower turn-off speeds due to increased gate capacitance, but it shows resilience under harsh conditions.

Limitations and Future Work

  • Limitations:
    • The paper focuses on a specific device structure (SSJ-UMOS).
  • While validated against TCAD and some experimental data, a broader experimental validation across various device designs would strengthen the findings.
  • The analytical model for dynamic characteristics is not as extensively detailed as the static characteristics.
  • Future Work:
    • The authors suggest that the proposed model can guide the design of power MOSFETs with SIPOS.
  • Further exploration of the impact of SIPOS pillar geometry and material properties on device performance could be beneficial.
  • Extending the machine learning approach to model other device characteristics or to explore a wider range of device architectures.

Summary updated

Abstract

This paper presents a machine learning-based figure of merit model for superjunction (SJ) U-MOSFET (SSJ-UMOS) with a modulated drift region utilizing semi-insulating poly-crystalline silicon (SIPOS) pillars. This SJ drift region modulation is achieved through SIPOS pillars beneath the trench gate, focusing on optimizing the tradeoff between breakdown voltage (BV) and specific ON-resistance (RON,sp). This analytical model considers the effects of electric field modulation, charge-coupling, and majority carrier accumulation due to additional SIPOS pillars. Gaussian process regression is employed for the figure of merit (FOM = BV2/RON,sp) prediction and hyperparameter optimization, ensuring a reasonable and accurate model. A methodology is devised to determine the optimal BV-RON,sp tradeoff, surpassing the SJ silicon limit. The paper also delves into a discussion of optimal structural parameters for drift region, oxide thickness, and electric field modulation coefficients within the analytical model. The validity of the proposed model is robustly confirmed through comprehensive verification against TCAD simulation results.
BibTeX
@inproceedings{cao2024machine,
  author = {Zhen Cao and Qi Sun and Chuanfeng Ma and Biao Hou and Licheng Jiao},
  title = {{Machine Learning-Based Figure of Merit Model of SIPOS Modulated Drift Region for U-MOSFET}},
  booktitle = {Micromachines},
  year = {2024},
  doi = {10.3390/mi15030411},
}
Figure