Title: PALTO: Physics-Informed Active Learning for Tri-Gate FinFET Design Optimization for Vertical Power Delivery

Abstract:

This study validates the utility of machine learning (ML)-based optimization in engineering application-specific GaN tri-gate FinFETs for vertical power delivery architectures. Traditional TCAD methods are often too computationally demanding and inadequate for managing the complex, high-dimensional, and nonlinear design landscapes characteristic of advanced GaN technology. To overcome these limitations, we employ a physics-informed active learning framework that intelligently directs simulations, thereby accelerating convergence without compromising precision. This ML-assisted strategy facilitates the identification of optimal device configurations by efficiently investigating critical structural variables, particularly the GaN-to-AlGaN thickness ratio, which has historically been a subject of significant debate in device engineering. Through a systematic exploration of these key parameters, two optimized devices featuring aggressively scaled gate-to-drain lengths were identified.

Single-fin, multi-channel simulations indicate that Device D2, characterized by a thinner GaN channel compared to the AlGaN barrier, yields a higher drive current. Conversely, in a 300-fin setup, Device D1 surpasses D2, delivering 3.3 A with an on-resistance of 0.49 ohm—a performance improvement of approximately 2$\times$—even though it exhibits slightly higher parasitic elements. Both devices function in a normally-off mode. When evaluated against an application-specific figure of merit, Device D1 achieves a value of 5 pC$\cdot$ohm, reflecting 2$\times$ superior switching efficiency compared to Device D2. Ultimately, both proposed designs demonstrate performance advantages over industrial benchmarks, albeit from different perspectives.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC