In a global first, researchers at Australia’s national science agency, CSIRO, have demonstrated that quantum machine learning (QML) can outperform classical artificial intelligence in semiconductor device modeling.
Highlights
- First Real-World Quantum AI Demo: CSIRO demonstrated the first practical use of quantum machine learning (QML) to outperform classical AI in semiconductor modeling.
- Core Focus — Ohmic Contact Resistance: The study targeted a crucial yet complex parameter in microchips that affects current flow, using real-world GaN HEMT device data.
- Hybrid Model Innovation: Researchers developed QKAR, a hybrid quantum-classical model using PCA, a five-qubit quantum kernel, and classical regression to model device behavior.
- Outperformed Classical Algorithms: QKAR beat seven traditional ML models, achieving a low MAE of 0.338 Ω·mm and maintaining accuracy under quantum noise conditions.
- From Prediction to Fabrication: Devices fabricated using QKAR’s predictions showed improved performance, validating its real-world applicability in chip manufacturing.
- Experimental First in Quantum AI: Unlike most quantum AI studies, this one used actual fabrication data instead of simulations, marking a major leap in applied quantum research.
- Scalable End-to-End Workflow: The pipeline included encoding, dimensionality reduction, quantum kernel learning, VAE-based augmentation, and physical device production.
- Quantum AI’s Practical Potential: The success of this project shows quantum computing can solve real-world engineering challenges, even with today’s imperfect hardware.
This development could mark a significant shift in how microchips are designed and optimized, especially in scenarios involving complex, small-data problems.
Targeting a Core Challenge in Microchip Design
The breakthrough focuses on modeling Ohmic contact resistance, a key parameter in semiconductors that influences how efficiently current flows between metal and semiconductor layers.
Accurately modeling this property is notoriously difficult, particularly when working with high-dimensional data and limited experimental samples.
To address this, the CSIRO team worked with 159 real-world samples of gallium nitride (GaN) high-electron-mobility transistors (HEMTs)—devices commonly used in high-frequency and high-power applications such as electric vehicles, 5G, and aerospace systems.
QKAR
The researchers, led by Professor Muhammad Usman, developed a hybrid quantum-classical model known as the Quantum Kernel-Aligned Regressor (QKAR). The approach involved:
- Dimensionality reduction: Reducing 37 fabrication variables to the five most critical via Principal Component Analysis (PCA)
- Quantum encoding: Mapping these five features to a quantum state using a Pauli-Z feature map on a five-qubit quantum device
- Kernel learning: Using a quantum kernel layer to identify subtle, nonlinear relationships
- Classical regression: Applying a Support Vector Regressor to generate performance predictions
Performance and Validation
In benchmark tests, QKAR surpassed seven leading classical machine learning algorithms, achieving a mean absolute error (MAE) of just 0.338 Ω·mm when predicting Ohmic contact resistance.
The model’s performance remained consistent even under simulated quantum noise, a key concern when deploying algorithms on today’s noisy intermediate-scale quantum (NISQ) hardware. This robustness reinforces QKAR’s practical viability using current-generation quantum systems.
From Prediction to Physical Chips
To assess the model’s real-world impact, the team fabricated new GaN devices based on QKAR’s recommendations.
These devices demonstrated improved performance, validating that the quantum-enhanced design strategy can generalize beyond its training data and translate into tangible manufacturing improvements.
First-of-Its-Kind Experimental Application in Quantum AI for Semiconductors
While many quantum AI models remain at the simulation or theoretical stage, this study stands out as the first experimental demonstration of QML applied directly to physical semiconductor data.
By working with actual GaN HEMT fabrication datasets, the team moved beyond simulation to prove that quantum-enhanced modeling can provide insights in applied R&D environments.
A Scalable End-to-End Quantum-Classical Pipeline
The project’s workflow illustrates a comprehensive and scalable hybrid pipeline, integrating both quantum and classical techniques:
- One-hot encoding of 37 fabrication parameters
- Dimensionality reduction using classical PCA
- Quantum feature mapping and kernel learning
- Classical regression via SVR
- Data augmentation using a Variational Autoencoder (VAE)
- Physical fabrication of optimized devices based on model outputs
Quantum AI Moves Beyond Theory
CSIRO’s work signals a meaningful step forward in the practical use of quantum computing for real-world problems. By extracting useful patterns from limited data, quantum models like QKAR show promise in addressing some of the toughest challenges in semiconductor engineering.
The success of this hybrid approach underscores that even current quantum systems—despite their hardware limitations—can contribute to next-generation microchip design and broader applications in materials science and electronics.
