Hybrid Classical-Quantum Neural Networks for Multi-Characteristic Co-Optimization of Recessed-Gate AlGaN/GaN MIS-HEMTs

This paper proposes a hybrid classical-quantum neural network (HQNN) that significantly outperforms classical baselines in optimizing six electrical characteristics of recessed-gate AlGaN/GaN MIS-HEMTs by leveraging experimental data, while demonstrating that circuit depth, parameter count, and specific entanglement strategies are critical for accuracy and near-term hardware viability.

The Gist

Imagine you are trying to bake the perfect cake, but instead of flour and sugar, your ingredients are invisible microscopic processes like "plasma treatment" and "chemical cleaning." You want the cake to taste just right (have the right electrical properties), but every time you bake one, it costs a fortune, and the oven behaves slightly differently each time.

This is the challenge faced by engineers making GaN transistors (tiny power switches used in electronics). They need to find the perfect recipe, but they can't afford to bake thousands of cakes to test every variation.

Here is how the authors of this paper solved that problem using a mix of old-school math and new-fangled "quantum" magic.

1. The Problem: The Expensive, Noisy Kitchen

In the real world, making these transistors is messy. Tiny changes in how they are cleaned or heated cause big changes in how they work.

• The Data Problem: You can't just simulate this on a computer perfectly because the real world is too chaotic. You have to actually build the chips to get data.
• The Cost: They only had data from 468 chips. In the world of Artificial Intelligence (AI), that's a tiny, almost non-existent dataset. Usually, AI needs millions of examples to learn well. With so few examples, standard AI models tend to "memorize" the noise rather than learning the actual rules, leading to bad predictions.

2. The Solution: A Hybrid "Quantum-Classical" Chef

The team built a new type of AI called a Hybrid Classical-Quantum Neural Network (HQNN). Think of it as a two-person cooking team:

• The Classical Chef (The Human): This part of the AI is like a standard computer. It takes the messy recipe instructions (24 different variables like temperature, time, and chemical types) and organizes them into a simple, easy-to-understand summary.
• The Quantum Sous-Chef (The Magic): This is the new part. It takes that summary and runs it through a "quantum circuit." Imagine this as a magical spice grinder that can mix flavors in ways a normal grinder can't. It uses the weird rules of quantum physics (like superposition and entanglement) to find hidden patterns in the data that the human chef missed.

3. How They Tested It

They didn't just guess which "quantum spice grinder" was best. They built 19 different designs (templates) and tested them all, like trying different shapes of cookie cutters to see which one makes the best cookies.

They found that:

• More complexity helps (up to a point): Circuits with more "knobs" to turn (parameters) and more layers of mixing (depth) worked better.
• The "Goldilocks" Zone: If the quantum circuit was too complex (too random), it actually got worse. It's like trying to mix a cake batter with a blender set to "maximum chaos"—you just get a mess. The best circuits were complex enough to find patterns but not so chaotic that they got lost.
• Better Tools: Circuits that used "adjustable" mixing tools (parameterized gates) worked better than those with "fixed" tools (static gates).

4. The Results: A Better Recipe

When they compared their new Hybrid Chef against a standard AI (the "Classical Baseline"), the Hybrid Chef won.

• The Score: It reduced the overall error by 24.4%.
• The Specific Wins:
  • It predicted the "on/off" switch behavior much better.
  • It was especially good at predicting the leakage (how much electricity leaks out when the switch is off). This is usually the hardest thing to predict because it's so sensitive to tiny manufacturing errors.
  • It predicted the "hysteresis" (how the memory of the switch changes) more accurately.

5. The "Noise" Test: Will it Work on Real Quantum Computers?

Real quantum computers today are "noisy"—they make mistakes, like a radio with static. The team simulated this noise to see if their model would break.

• The Finding: Even with a moderate amount of "static" (noise), the model still worked very well. It only started to struggle when the noise was extremely high.
• The Takeaway: This suggests that we don't need a perfect, futuristic quantum computer to use this method. We could potentially run this on the small, imperfect quantum computers available right now.

Summary

The paper shows that by combining a standard computer with a small, specialized quantum circuit, engineers can learn the "secret recipes" for making better transistors, even when they only have a tiny amount of expensive data. It's like using a magic lens to see patterns in a blurry photo that a normal eye would miss, helping them build better electronics faster and cheaper.

Technical Summary

Technical Summary: Hybrid Classical-Quantum Neural Networks for Multi-Characteristic Co-Optimization of Recessed-Gate AlGaN/GaN MIS-HEMTs

Problem Statement
Optimizing recessed-gate AlGaN/GaN Metal-Insulator-Semiconductor High Electron Mobility Transistors (MIS-HEMTs) requires accurate modeling of multiple coupled electrical characteristics. These devices are sensitive to process-induced variability (e.g., recess depth, dielectric composition, surface treatment) that analytical models cannot capture without sacrificing accuracy or efficiency. While Machine Learning (ML) has become a standard for semiconductor modeling, classical approaches often struggle with experimental datasets. These datasets are costly to generate, inherently small, and noisy due to fabrication stochasticity. Dense classical networks frequently overfit on such data or fail to capture complex, non-linear correlations between process variables and multiple coupled device targets (threshold voltage, hysteresis, subthreshold swing, on/off currents) simultaneously.

Methodology
The authors propose a Hybrid Classical-Quantum Neural Network (HQNN) framework designed for joint optimization of six electrical targets from a 24-dimensional fabrication/process vector. The methodology involves:

1. Dataset: The study utilizes 468 experimentally fabricated devices across 17 distinct process splits. The input vector includes continuous variables (wafer coordinates, geometry) and binary one-hot encoded categorical process choices (cleaning, plasma treatment, dielectric deposition, annealing). Six targets are predicted: forward/reverse threshold voltages ($V_{th,fwd}$, $V_{th,rev}$), hysteresis ($\Delta V_{th}$), subthreshold swing (SS), on-current ($I_{on}$), and off-current ($I_{off}$).
2. Architecture: The HQNN employs a "strict-bottleneck" design where a classical Multi-Layer Perceptron (MLP) encoder compresses the 24-dimensional input into a 4-dimensional latent code. This code is mapped to rotation angles for a 4-qubit Parameterized Quantum Circuit (PQC). The PQC processes the state through variational layers, and a 3-axis Pauli measurement (X, Y, Z) yields a 12-dimensional feature vector, which is linearly mapped to the six targets.
3. Circuit Design & Screening: The authors systematically screened a family of 19 quantum circuit templates derived from Sim et al. [18], varying connectivity (nearest-neighbor, all-to-all), entangling gate types (CNOT, CRZ, CRX), and circuit depth ($L$). They evaluated both "single-template" (repeating one block) and "sequential mixed-circuit" (stacking different blocks) ansatzes.
4. Training: The model is trained using a weighted multi-target Mean Squared Error (MSE) loss. Gradients for the quantum parameters are computed via the parameter-shift rule, integrated with classical backpropagation through the encoder and readout layers using CUDA-Q and PyTorch.

Key Contributions

• Systematic Circuit Screening: The work provides a large-scale ablation study of quantum circuit templates for semiconductor regression, identifying specific architectural properties that correlate with performance.
• Multi-Target Co-Optimization: Unlike prior quantum kernel methods that often rely on PCA compression or single-target regression, this HQNN performs end-to-end supervised learning on the full feature vector to predict multiple coupled device metrics simultaneously.
• Design Guidance: The study establishes correlations between circuit properties (expressibility, parameter count, depth, entangler type) and prediction accuracy, offering a framework for selecting ansatzes for small-data problems.

Results
The selected HQNN, identified as Circuit (13, 5) at $L=2$ (a mixed-circuit configuration using template 13 at level 1 and template 5 at level 2), was benchmarked against classical baselines including ANN, PLS, Decision Trees, XGBoost, and others.

• Performance Gain: The HQNN reduced the overall normalized Root Mean Square Error (nRMSE) by 24.4% compared to the best classical baseline (ANN).
• Target-Specific Improvements:
  • $V_{th,fwd}$ RMSE: 0.297 V $\to$ 0.270 V
  • $V_{th,rev}$ RMSE: 0.278 V $\to$ 0.263 V
  • $\Delta V_{th}$ RMSE: 0.049 V $\to$ 0.045 V
  • SS RMSE: 22.22 mV/dec $\to$ 19.87 mV/dec
  • $I_{off}$ RMSE: $5.75 \times 10^{-8}$ A $\to$ $4.35 \times 10^{-8}$ A (a ~24% reduction)
  • $I_{on}$ RMSE remained competitive (0.053 A vs. 0.056 A for ANN).
• Ablation Insights:
  • Expressibility: Accuracy correlated negatively with expressibility ($D_{KL}$). Highly expressive (Haar-random-like) circuits performed worse, likely due to barren plateaus in small-data regimes.
  • Parameters & Depth: Accuracy correlated positively with parameter count, circuit depth, and two-qubit gate count.
  • Entanglers: Parameterized controlled-rotation gates (CRZ, CRX) outperformed static CNOT-based circuits.
• Noise Robustness: A depolarizing noise study on a representative circuit (Template 13, $L=4$) showed that the model maintains an overall $R^2$ of 0.905 under mild noise ($p=0.005$). While $I_{off}$ prediction was sensitive to noise, voltage-related targets remained stable, suggesting the classical encoder retains predictive power even if the quantum representation degrades.

Significance and Claims
The paper claims to advance Quantum Machine Learning (QML) in semiconductor modeling from single-target quantum kernel regression toward multi-characteristic device modeling for recessed-gate GaN MIS-HEMTs. The authors assert that variational hybrid quantum models can improve generalization on small, noisy experimental datasets where classical ML is constrained.

The work suggests that comparable HQNNs may be trainable or deployable on near-term quantum hardware (specifically 4-qubit registers) given their demonstrated robustness to mild depolarizing noise. The study provides a circuit-design framework for future QML-assisted semiconductor process optimization, emphasizing that controlled ansatz selection is critical for realizing quantum advantages in practical, data-constrained engineering tasks.