⚡ electrical engineering

Physics-Based Performance Mapping of Ultra-High-Voltage SiC IGBTs and MOSFETs With Machine-Learning Acceleration

This paper presents a physics-based, machine-learning-accelerated electrothermal framework that efficiently compares the performance of ultra-high-voltage SiC IGBTs, SiC MOSFETs, and their series-connected silicon counterparts under hard and soft switching, revealing that while SiC IGBTs excel in low-to-medium frequency zero-voltage switching, series-connected SiC MOSFETs offer superior performance under hard switching conditions.

Original authors: Nazareno Donato

Published 2026-10-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nazareno Donato

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

High-voltage power electronics are the silent workhorses of the modern grid, responsible for moving massive amounts of electricity from power plants to our homes and factories. To do this efficiently, engineers rely on switches that can turn current on and off billions of times a second without wasting energy as heat. For decades, silicon has been the material of choice for these switches, but as demands for higher voltages and greater efficiency grow, scientists are turning to a harder, more resilient material called silicon carbide. Within this material, two main types of electronic switches compete: the MOSFET, which is fast and simple, and the IGBT, which is designed to handle heavy currents more easily. The question facing engineers is not just which one works, but which one is the best choice for a specific job, a decision that depends on a complex mix of voltage, temperature, and how often the switch flips.

Researchers at the University of Cambridge have tackled this question by creating a detailed digital map that predicts which switch will perform best under nearly any condition. Instead of building physical prototypes for every possible scenario, which would be slow and expensive, they used advanced computer simulations to model the physics of these devices. They focused on ultra-high-voltage applications, specifically looking at devices capable of handling 20,000 volts, a level of power that pushes the limits of current technology. The team compared single silicon carbide switches against two other strategies: using multiple silicon carbide switches connected in a line, and using multiple older silicon switches connected in a line. To make sense of the vast amount of data generated by these simulations, they employed a machine-learning tool, a type of computer program that learns patterns from data, to speed up the process of finding the best solution.

The core of their work involved simulating how these devices behave when they are subjected to intense electrical stress and heat. They tested the switches under two main conditions: "hard switching," where the device turns on or off while voltage is still present, and "soft switching," where the voltage is zero at the moment of switching, which is gentler on the components. The simulations revealed that the best choice is not a single winner but depends heavily on the operating environment. For the 20,000-volt class, the silicon carbide IGBT proved to be the strongest contender when the system operated at low to medium speeds and used the gentler soft-switching method. In these scenarios, the IGBT could handle nearly twice the current of a standard silicon carbide MOSFET before overheating. This advantage comes from the IGBT's ability to conduct electricity more efficiently when the current is high, a trait that becomes less useful as the switching speed increases.

However, the landscape changes when the switches are forced to work harder or faster. When the researchers included the energy losses that occur during hard switching, the advantage of the IGBT shrank significantly. In these tougher conditions, and especially when the material inside the IGBT had a shorter lifespan for its charge carriers, the series-connected silicon carbide MOSFET configuration emerged as the superior choice. This setup, which links three lower-voltage MOSFETs together to handle the high voltage, maintained its performance better at higher frequencies. The study also looked at connecting older silicon IGBTs in a series, but the simulations showed this approach generally performed the worst, struggling to compete with the silicon carbide options across the board.

To ensure these findings were reliable without taking years to compute, the researchers developed a hybrid approach that combined precise physics calculations with machine learning. They first ran thousands of detailed simulations to create a training dataset, teaching the computer how the different switches behave. Then, they used the machine-learning model to predict the best device for most operating points, only running the slow, detailed physics check when the computer was unsure or when the results were very close. This method sped up the creation of the performance maps by more than two times, allowing the team to explore a vast range of possibilities quickly. The resulting maps show clear boundaries where one technology should be chosen over another, helping engineers design more efficient power systems without needing to guess.

The study concludes that while silicon carbide IGBTs are excellent for high-current, low-frequency applications, particularly when the system is designed to switch gently, they are not a universal solution. For high-frequency operations or systems that must switch rapidly, the series-connected silicon carbide MOSFET offers a more robust path forward. The research does not claim to have solved all challenges in high-voltage power electronics, but it provides a powerful, physics-based tool for screening technologies early in the design process. By mapping out exactly where each device excels, the work helps engineers make informed decisions before they commit to building expensive hardware, ensuring that the next generation of power grids is both efficient and reliable.

Technical Summary: Physics-Based Performance Mapping of Ultra-High-Voltage SiC IGBTs and MOSFETs With Machine-Learning Acceleration

Problem Statement
The selection of optimal power devices for ultra-high-voltage (UHV) applications (e.g., 20 kV class) is complicated by the trade-offs between emerging Silicon Carbide (SiC) Insulated-Gate Bipolar Transistors (IGBTs), SiC MOSFETs, and series-connected configurations of lower-voltage Si IGBTs and SiC MOSFETs. While SiC IGBTs offer conductivity modulation to reduce conduction losses, their commercial adoption is hindered by challenges in carrier-lifetime engineering and bipolar degradation. Conversely, series-connected devices introduce parasitic inductances and control complexity. Traditional compact models often fail to capture critical physical effects in SiC, such as incomplete ionization, anisotropic transport, and carrier recombination, necessitating computationally expensive Physics-Based Finite Element Method (FEM) simulations. Furthermore, generating comprehensive performance maps across wide ranges of current, switching frequency, duty cycle, and temperature using full electrothermal solvers is prohibitively slow for rapid technology screening.

Methodology
The author proposes a self-consistent electrothermal framework accelerated by machine learning (ML) to compare four device architectures: monolithic SiC MOSFETs, SiC IGBTs, series-connected SiC MOSFETs, and series-connected Si IGBTs.

TCAD Modeling: Mixed-mode FEM simulations were performed on active cell structures for 1.2 kV to 20 kV voltage classes. The models incorporated detailed physics, including incomplete ionization, Auger/SRH/radiative recombination, and temperature-dependent carrier lifetimes. Switching losses were extracted via double-pulse test (DPT) circuits, distinguishing between hard switching and zero-voltage switching (ZVS) conditions.
Electrothermal Solver: A deterministic solver calculates total power loss ( $P_{tot}$ ) as the sum of conduction and switching losses. It iteratively solves for the self-consistent junction temperature ( $T_j$ ) using a fixed-point iteration method until convergence, ensuring the device remains within thermal limits ( $T_{j,max} = 448$ K).
Three-Stage Analysis:
- Maximum-Current Capability ( $I^*$ ): Determines the thermal current envelope for each device at specific operating points.
- Deterministic Superior-Device Mapping: Ranks devices at fixed operating points based on minimum total loss, identifying the "winner" and the loss margin ( $\Delta P$ ) against the runner-up.
- ML-Accelerated Mapping: To address the computational cost of sweeping dense grids ( $N_{op} \approx 61,875$ points), a hybrid approach was developed. A bagged decision-tree ensemble (trained on a subset of deterministic solutions) predicts the optimal device. A "deterministic correction" step re-evaluates points where the ML confidence is low (below a threshold of 0.88) or near class boundaries.

Key Contributions

Unified Comparison Framework: The study provides a rigorous, physics-based comparison of UHV SiC IGBTs, SiC MOSFETs, and their series-connected counterparts under both hard and soft-switching conditions, accounting for thermal constraints and carrier lifetime variations.
Hybrid ML Acceleration: The paper introduces a specific ML workflow that accelerates the generation of device-selection maps by 2.1× (excluding training time) compared to a full deterministic sweep. Crucially, this approach does not replace the physics solver but uses it to correct predictions in transition regions, ensuring accuracy where device performance is comparable.
Carrier Lifetime Sensitivity: The framework explicitly models the impact of carrier lifetime ( $\tau$ ) on SiC IGBT performance, demonstrating how reduced lifetimes degrade the conductivity modulation advantage.

Results

20 kV Class Performance: Under ZVS (soft switching) conditions, the SiC IGBT demonstrates superior current capability in the low-to-medium frequency range (approx. 100 Hz to 20 kHz), supporting nearly double the current of a SiC MOSFET at a 50% duty cycle. This advantage is driven by lower conduction losses.
Impact of Hard Switching: When turn-on losses are included (hard switching), the crossover frequency shifts significantly (below 1 kHz), and the SiC IGBT's current capability drops. For shorter carrier lifetimes ( $\tau = 0.1 \mu s$ ), the SiC IGBT loses its advantage, and the series-connected SiC MOSFET configuration becomes superior, particularly at higher frequencies and lower currents.
Series-Connected Devices: Series-connected Si IGBTs showed the poorest performance across the investigated space, remaining below 30 A. Series-connected SiC MOSFETs offered a competitive alternative to monolithic SiC IGBTs in high-frequency, low-current regimes.
6.5 kV Class: At lower voltage classes, the SiC IGBT advantage largely disappears; the SiC MOSFET outperforms the IGBT across the entire frequency range due to the high built-in voltage of SiC bipolar junctions.
Computational Efficiency: The hybrid ML approach reduced the runtime for a full map generation from ~179 seconds (deterministic only) to ~85 seconds, achieving a 2.1× speed-up while maintaining accuracy in transition boundaries.

Significance and Claims
The paper claims that this framework serves as an efficient "first-stage" technology screening method. It allows researchers and engineers to rapidly evaluate the trade-offs between different UHV device architectures over wide parameter ranges before committing to detailed compact-model extraction or experimental converter-level verification. The author modestly states that while the ML acceleration provides moderate gains for the current grid density, the benefits would scale significantly for larger device libraries, denser grids, or more complex parasitic configurations. The study concludes that the optimal technology is highly dependent on the specific operating point (current, frequency, duty cycle), thermal conditions, and the achievable carrier lifetime in SiC IGBTs, with no single device being universally superior across all UHV scenarios.

More like this