Source-linked AI summary
A Unified Physics-Aware Quantum Machine Learning Framework across Power GaN HEMTs and Logic Nanowire FETs: Predicting Unseen Process Splits and Held-Out Geometry Combinations with Lower Error and Tighter Split-to-Split Variability
Rushat Rai, Yun-Yuan Wang, Autsada Kakaen, Pei-Jie Chang, Doan Viet Nguyen, Yuan-Chieh Chiu, Doldet Tantraviwat, Niall Tumilty, Simon See, Wen-Jay Lee, Tai-Yue Li, Nan-Yow Chen, Tian-Li Wu
TL;DR
Device modeling for unseen process splits and geometry combinations is constrained by costly fabrication and TCAD cycles. This paper uses reinforcement learning to search compact quantum circuits for OOD device prediction, achieving the lowest MAE across all 11 targets against six classical baselines, with lower error and tighter fold variability on HEMTs and NWFETs.
Problem
Device optimization for unseen process splits and geometries is limited by costly fabrication and TCAD cycles and stochastic device variation.
Method
A GNN policy trained with PPO searches compact PQC architectures using nested LOGOCV error on held-out process or geometry groups for OOD device modeling.
Results
The RL-selected PQC achieves the lowest OOD MAE on all 11 targets versus six classical baselines, with reductions reaching 59% for HEMTs and 84% for NWFETs.
Takeaways & Limitations
The framework provides a compact surrogate that generalizes across measured HEMT process splits and unseen NWFET geometry combinations without device-specific equations.
Abstract
from arXiv · showhide
We present a unified reinforcement-learning (RL) framework that discovers compact parametrized quantum circuits (PQCs) for data-scarce device modeling. A graph neural network (GNN) policy optimized by proximal policy optimization (PPO) searches circuit architectures using leave-one-group-out cross-validation (LOGOCV) error on held-out process or geometry groups as the reward. The framework achieves the lowest mean absolute error (MAE) on all 11 targets versus six classical baselines, with 59% lower error (Ioff) and 81% tighter fold variability (VTH) for HEMTs and 84% lower error (VTH, SS, Ioff) and 82% tighter fold variability (Ioff) for NWFETs. These results demonstrate the potential of RL-selected, classically simulated PQCs as compact surrogates with low OOD error and improved physical consistency, despite imposing no explicit physical constraints, penalty terms, or device-specific equations, on the two evaluated device datasets.
I. INTRODUCTION
The section motivates unified modeling for power GaN HEMTs and sub-3-nm logic NWFETs, and presents an experimentally validated RL architecture-searched PQC framework optimized for multi-target OOD prediction.
- p-GaN gate GaN HEMTs support normally-off operation with controllable threshold voltage for high-performance power switching.
- Gate-all-around NWFETs suppress short-channel effects and off-state leakage for continued scaling toward low-power logic.
- The work presents the first experimentally validated RL architecture-searched PQC framework for multi-target device modeling optimized directly for OOD prediction.
II. METHODOLOGY
The methodology combines physics-based HEMT pretraining and staged experimental adaptation with direct single-stage TCAD training for NWFETs. Nested LOGOCV selects robust PQCs, while hybrid quantum-classical training uses mean-squared error without physics-informed losses or constraint terms.
- Dataset preparation: A 1D electrostatic solver models Mg-acceptor ionization, polarization charge, 2DEG formation, and interface traps for synthetic p-GaN HEMT pretraining.Finite-volume discretization with damped Newton iteration extracts full transfer curves.
- Dataset preparation: HEMT training proceeds through synthetic electrostatic pretraining, frozen-PQC wafer-position-head calibration, and joint experimental fine-tuning.NWFETs use direct single-stage TCAD training.
- Architecture selection: Nested LOGOCV rewards the worst validation-group error while applying complexity and invalid-structure penalties, then evaluates the selected candidate on an untouched outer group.This prevents strong average performance from concealing failure on one split.
- Hybrid training: Hybrid quantum-classical training minimizes mean-squared error alone, using parameter-shift gradients on NVIDIA CUDA-Q without physics-informed loss or constraint terms.The training objective therefore does not impose explicit physics-based penalties or constraints.
III. RESULTS AND DISCUSSION
The RL-selected hybrid quantum neural network is evaluated against six classical baselines on process-variation and geometry-scaling OOD challenges using leakage-resistant group-wise validation. For NWFET joint-geometry generalization, it substantially reduces out-of-fold MAE versus ANN across three radius-sensitive metrics.
- Evaluation protocol: The PQC is compared with six classical baselines on measured p-GaN HEMT process variation and 3D NWFET geometry scaling.All models use the same group-wise validation protocol, with results averaged over 5 random seeds.
- Evaluation protocol: The PQC exceeds ANN in sign retention and trend agreement across all LOGO folds, with lower fine-tuning-induced degradation.The comparison uses the same data pipeline, wafer-position correction, and training stages.
- Joint-geometry OOD generalization in NWFETs: 50.1%, 46.3%, and 39.2%: PQC out-of-fold MAE reductions versus ANN for NWFET off-current, subthreshold swing, and maximum transconductance, respectively.The joint-geometry OOD evaluation covers electrostatic gate control across subthreshold and on-state regimes.
IV. CONCLUSION
The paper presents an experimentally validated reinforcement-learning-searched PQC framework for out-of-distribution device prediction. It unifies multi-target modeling across p-GaN HEMTs and 3D-NWFETs without embedding device-specific equations.
- The framework combines multi-target OOD prediction, RL-searched quantum circuits, and scalability across power and logic device families.It generalizes across measured p-GaN HEMT process splits and unseen 3D-NWFET geometry combinations.
- A common data-reuploading PQC workflow adapts across devices without re-deriving partial differential equations.The framework is designed for physically consistent OOD extrapolation in emerging-device surrogate modeling.
- A relational graph-attention GNN policy searches PQC architectures by predicting the next circuit operation from device-feature inputs.The search space includes rotation and entanglement blocks represented in a heterogeneous graph of feature, qubit, block, and readout nodes.
- Nested LOGOCV supplies the reinforcement-learning reward while solver-informed pretraining and experimental fine-tuning support OOD evaluation.For NWFETs, the workflow uses direct TCAD-only training because solver-pretraining data are unavailable.