News

Quantum Computing Shows Promise for Semiconductor Manufacturing

Researchers develop quantum machine learning model that outperforms classical methods in predicting chip component performance

• By John K. Waters
• 08/01/2025

Scientists have demonstrated that quantum machine learning can outperform traditional computer models in predicting the performance of critical semiconductor components, potentially offering manufacturers a new tool for optimizing chip production processes.

A research team developed a quantum kernel-aligned regressor (QKAR) that successfully predicted the electrical resistance of ohmic contacts in gallium nitride transistors using only 159 experimental samples. The quantum model consistently achieved lower error rates than seven classical machine learning approaches tested in the same study.

The work, published in Advanced Science, an academic journal, represents the first application of quantum machine learning to modeling ohmic contact formation in high-electron-mobility transistors, components essential for high-power and high-frequency electronics used in everything from cell phone amplifiers to electric vehicle power systems.

Semiconductor manufacturing faces a persistent challenge: the high cost and time required to collect experimental data makes it difficult to build robust predictive models. Traditional machine learning methods typically require large datasets to perform well, but gathering sufficient data from semiconductor fabrication processes can take months and cost hundreds of thousands of dollars.

"Classical machine learning approaches have been successful in many domains, but their performance degrades in small-sample, nonlinear scenarios," the researchers wrote. The quantum approach exploits quantum kernels to capture intricate correlations from compact datasets.

The QKAR model achieved a mean absolute error of 0.338 ohm-millimeters when predicting contact resistance, representing at least a 15 percent improvement over the best classical model tested. The quantum system showed particular strength in handling the complex, nonlinear relationships between manufacturing parameters such as annealing temperature, metal composition, and atmospheric conditions.

The researchers used a five-qubit quantum circuit combining a Pauli-Z feature map with a trainable quantum kernel alignment layer. Unlike more complex quantum neural networks, this relatively simple architecture proved most effective for the semiconductor modeling task.

"While increasing the entanglement depth in quantum circuits is often expected to improve performance, our results suggest that deeper embeddings may degrade performance with limited dataset sizes," the team found. More complex quantum circuits can transform data into quantum states that are too fragmented, making it harder to detect meaningful patterns.

To validate their model, researchers fabricated five additional test devices with different material compositions and processing conditions. The quantum model maintained its accuracy across these variations, suggesting it could generalize to new manufacturing scenarios.

Current quantum computers are inherently noisy, raising questions about whether quantum machine learning advantages would persist on actual hardware. The researchers tested their model under simulated noise conditions equivalent to those found in near-term quantum processors.

With single-qubit error rates up to 1%, the quantum model maintained its performance advantage over classical methods. However, at unrealistically high noise levels of 5%, the system failed to function properly.

"Most state-of-the-art quantum platforms exhibit effective error rates below 0.5 percent," the researchers noted. Combined with ongoing improvements in quantum error correction, the results suggest quantum machine learning could be viable for real-world semiconductor applications in the near future.

The semiconductor industry has increasingly turned to artificial intelligence and machine learning to optimize manufacturing processes as chip designs become more complex. However, the industry's reliance on expensive experimental validation has limited the effectiveness of data-hungry classical approaches.

Quantum machine learning could offer particular advantages for specialized manufacturing processes where data collection is especially costly or time-consuming. The ability to extract meaningful patterns from small datasets could help manufacturers optimize new processes more quickly and cost-effectively.

The research team acknowledged that classical machine learning models might achieve better performance with more extensive optimization. They used standard configurations for comparison rather than exhaustively tuning parameters for each method.

Beyond ohmic contacts, the quantum approach could potentially apply to other semiconductor manufacturing challenges where process parameters interact in complex, nonlinear ways. These might include metal deposition, etching processes, or the formation of other critical device components.

The quantum circuits used in the study are compatible with current near-term quantum processors, suggesting the approach could transition from simulation to real quantum hardware as the technology matures.

As quantum processors improve in reliability and scale, quantum machine learning may become a standard tool in semiconductor research and development, offering manufacturers new ways to optimize processes with limited experimental data.

Australia's CSIRO supported the research and represents ongoing efforts to find practical applications for quantum computing in industrial processes.