Work

Ψ-HDL: Physics Structure-Informed Neural Networks for Hardware Description Language Generation

Year: 2026
Type: article
Abstract: Compact device modeling remains a critical bottleneck in hardware innovation: traditional physics-based derivation demands weeks or months of expert effort, while closed-box machine learning models cannot be deployed in circuit simulators. We present $\Psi $ -HDL (Physics Structure-Informed Neural Networks for Hardware Description Language Generation), an automated framework that bridges this gap by learning interpretable, physically consistent models and generating synthesizable Verilog-A code within minutes. $\Psi $ -HDL integrates three breakthrough ideas: 1) physics-informed neural networks that embed governing equations into training, ensuring models respect conservation laws and passivity constraints; 2) knowledge distillation with relation-aware regularization that compresses networks by 97.6-99.8% while exposing weight symmetries and parameter sharing patterns; and 3) direct HDL translation that maps discovered structures into human-readable, circuit-ready equations. The framework automatically uncovers physical insights, including odd/even function symmetries, voltage-state coupling modes, and transport mechanisms, all without manual specification. We validate $\Psi $ -HDL across four demanding domains: Burgers and Laplace equations achieve $54\times $ and $2.3\times $ accuracy improvements over standard PINNs; a neuromorphic XOR circuit achieves 100% logic fidelity with a minimal 6-memristor topology ( $120~\mu $ m ${}^{2}$ ); and most critically, on the memristor case study we observe an explicit accuracy–compression trade-off: polynomial regression and LUT interpolation achieve lower interpolation error on the clean synthetic dataset, while $\Psi $ -HDL targets physics constraints, interpretability, and simulator-deployable compact representations. The physics-informed teacher achieves MAE $= 1.09\times 10.4$ A and improves upon the VTEAM baseline by 28.7%, whereas the compressed HDL-ready model uses 13 cluster centers (96 tied parameters in the reconstructed network) with MAE $= 3.645\times 10.4$ A. A comprehensive analysis across three device physics (oxide, phase-change, and organic) and seven network architectures (20-500 neurons) demonstrates robust generalization. $\Psi $ -HDL generates Verilog-A models targeting commercial simulators (e.g., Cadence Spectre). For ngspice-based validation, we generate an equivalent behavioral SPICE subcircuit (B-sources) and benchmark simulator-level agreement against the PyTorch reference. By embedding learned physical structure in compact HDL code, $\Psi $ -HDL eliminates the manual modeling bottleneck, accelerating the integration of emerging devices from weeks to minutes while preserving the interpretability essential for trustworthy circuit design. All code, models, and validation data are open-source.
Source: IEEE Access
Language: English
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7128301720
FWCI: 3.6
Cites: 26
Cited by: 3
Related to: 10