Modeling MOSFET behavior using automatic differentiation

Scientists from Nara Institute of Science and Technology (NAIST) utilized the mathematical method called automatic differentiation to find the optimal acceptable of experimental information up to 4 times faster. This probe tin beryllium applied to multivariable models of physics devices, which whitethorn let them to beryllium designed with accrued show portion consuming little power.

Wide bandgap devices, specified arsenic silicon carbide (SiC) metal-oxide semiconductor field-effect transistors (MOSFET), are a captious constituent for making converters faster and much sustainable. This is due to the fact that of their larger switching frequencies with smaller vigor losses nether a wide scope of temperatures erstwhile compared with accepted silicon-based devices. However, calculating the parameters that find however the electrical current successful a MOSFET responds arsenic a relation of the applied voltage remains hard successful a circuit simulation. A amended attack for fitting experimental information to extract the important parameters would supply spot manufacturers the quality to plan much businesslike powerfulness converters.

Now, a squad of scientists led by NAIST has successfully utilized the mathematical method called automatic differentiation (AD) to importantly accelerate these calculations. While AD has been utilized extensively erstwhile grooming artificial neural networks, the existent task extends its exertion into the country of model parameter extraction. For problems involving galore variables, the task of minimizing the mistake is often accomplished by a process of "gradient descent," successful which an archetypal conjecture is repeatedly refined by making tiny adjustments successful the absorption that reduces the mistake the quickest. This is wherever AD tin beryllium overmuch faster than erstwhile alternatives, specified arsenic symbolic oregon numerical differentiation, astatine uncovering absorption with the steepest "slope". AD breaks down the occupation into combinations of basal arithmetic operations, each of which lone needs to beryllium done once. "With AD, the partial derivatives with respect to each of the input parameters are obtained simultaneously, truthful determination is nary request to repetition the exemplary valuation for each parameter," archetypal writer Michihiro Shintani says. By contrast, symbolic differentiation provides nonstop solutions, but uses a ample magnitude of clip and computational resources arsenic the occupation becomes much complex.

To amusement the effectiveness of this method, the squad applied it to experimental information collected from a commercially disposable SiC MOSFET. "Our attack reduced the computation clip by 3.5× successful examination to the accepted numerical-differentiation method, which is adjacent to the maximum betterment theoretically possible," Shintani says. This method tin beryllium readily applied successful galore different areas of probe involving aggregate variables, since it preserves the carnal meanings of the exemplary parameters. The exertion of AD for the enhanced extraction of exemplary parameters volition enactment caller advances successful MOSFET improvement and improved manufacturing yields.

The probe was published successful IEEE Transactions connected Power Electronics.

---

---

More information: Michihiro Shintani et al, Accelerating Parameter Extraction of Power MOSFET Models Using Automatic Differentiation, IEEE Transactions connected Power Electronics (2021). DOI: 10.1109/TPEL.2021.3118057

Provided by Nara Institute of Science and Technology