Zero-Server In-Browser Adjoint Sensitivity Analysis and Analog Sizing via WebGPU and Mixed-Precision Checkpointing
Year: 2026
Type: preprint
Abstract: Traditional analog sizing and inverse design typically necessitate numerous forward SPICE simulations or rely on heavy, workstation-class adjoint and differentiable engines. While browser-native Electronic Design Automation (EDA), including prior KING TSENG AI Web EDA frameworks and WebAssembly (WASM)-based ngspice tools, has successfully brought forward simulation and Physics-Informed Neural Network (PINN) surrogates to the client, it has fundamentally lacked a native reverse-mode sensitivity path. This paper presents a zero-server adjoint sensitivity architecture for implicit transient and DC simulations. Executed entirely within the browser utilizing the WebGPU sparse solver ecosystem, the proposed method leverages a checkpointed backward recurrence tailored to accommodate browser memory constraints and the absence of ubiquitous 64-bit floating-point (fp64) arithmetic. This methodology yields parameter gradients of user-defined figures of merit at a computational cost scaling by only a small constant factor relative to a single forward transient solve. Consequently, projected-gradient and Adam-based device sizing can be performed entirely client-side. We establish a validation protocol against closed-form RC sensitivities, ngspice finite-difference gradients, and a two-stage amplifier sizing task. Positioned as a tool for exploration, pedagogy, and pre-layout "what-if" analysis rather than a foundry sign-off replacement, this framework bridges the gap between interactive web-based EDA and modern differentiable design paradigms.
Author: TSENG, I-WEI
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236106330
