18.1 A Self-Health-Learning GaN Power Converter Using On-Die Logarithm-Based Analog SGD Supervised Learning and Online Tj-Independent Precursor Measurement

Abstract

As GaN technology proliferates in modern power electronics, reliability of GaN-based circuits has become the biggest hurdle for commercialization. Sustaining largest voltage and current stresses in power circuits, power devices on average account for over 31% of failures [1]. With new problems such as current collapse and thermal aging, GaN power circuits deem to face more reliability challenges compared to their silicon counterparts [2]. In such a situation, health condition monitoring is of paramount importance. As shown in Fig. 18.1.1, due to hot electron injection and charge trapping effects, current collapse weakens 2-dimensional electron gas (2DEG) layer in a GaN switch over time, elevating its dynamic on-resistance $\mathrm{r}_{\mathrm{DS}_{-}\mathrm{On}}$ gradually. The clear link between $\mathrm{r}_{\mathrm{DS}_{-}\mathrm{ON}}$ and aging (Fig. 18.1.1) makes $\mathrm{r}_{\mathrm{DS}_{-}\mathrm{ON}}$ a widely accepted precursor for GaN condition monitoring [3]–[5]. However, measuring rDS_ ON is not a simple task. Traditionally, r DS_ ON can be measured offline by shutting- down the affiliated circuit. However, the- approach can be highly inaccurate due to significant discrepancy between offline and online operation conditions. To mitigate this issue, in-situ condition monitoring can be employed [3], [4]. However, it still requires designated test periods, causing interruptions of operation and increased test cost. A recent study applies machine learning (ML) to achieve online aging prognosis [5]. However, the ML algorithm is generic and is built on a standard digital basis. It requires sophisticated data processing and communication modules, causing substantial power and cost overheads. More importantly, the off-board look-up-table-based training process has to be performed offline, leading to similar drawbacks encountered in other approaches. Overall, all approaches reviewed here demand significant resources and time for either trimming, calibration or training in order to compensate for variations and errors induced by the fabrication process, work condition, user influence, etc. It would be much more desirable and efficient if a “plug-and-play” online aging prognosis method can be developed, which, as an essential part of a power circuit, requires no trimming and calibration.