Researchers at Australia’s CSIRO have achieved a world-first demonstration of quantum machine learning in semiconductor fabrication. The quantum-enhanced model outperformed conventional AI methods and could reshape how microchips are designed. The team focused on modeling a crucial—but hard to predict—property called “Ohmic contact” resistance, which measures how easily current flows where metal meets a semiconductor.
They analysed 159 experimental samples from advanced gallium nitride (GaN) transistors (known for high power/high-frequency performance). By combining a quantum processing layer with a final classical regression step, the model extracted subtle patterns that traditional approaches had missed.
Tackling a difficult design problem
According to the study, the CSIRO researchers first encoded many fabrication variables (like gas mixtures and annealing times) per device and used principal component analysis (PCA) to shrink 37 parameters down to the five most important ones. Professor Muhammad Usman – who led the study – explains they did this because “the quantum computers that we currently have very limited capabilities”.
Classical machine learning, by contrast, can struggle when data are scarce or relationships are nonlinear. By focusing on these key variables, the team made the problem manageable for today’s quantum hardware.
A quantum kernel approach
To model the data, the team built a custom Quantum Kernel-Aligned Regressor (QKAR) architecture. Each sample’s five key parameters were mapped into a five-qubit quantum state (using a Pauli-Z feature map), enabling a quantum kernel layer to capture complex correlations.
The output of this quantum layer was then fed into a standard learning algorithm that identified which manufacturing parameters mattered most. As Usman says, this combined quantum–classical model pinpoints which fabrication steps to tune for optimal device performance.
In tests, the QKAR model beat seven top classical algorithms on the same task. It required only five qubits, making it feasible on today’s quantum machines. CSIRO’s Dr. Zeheng Wang notes that the quantum method found patterns classical models might miss in high-dimensional, small-data problems.
To validate the approach, the team fabricated new GaN devices using the model’s guidance; these chips showed improved performance. This confirmed that the quantum-assisted design generalized beyond its training data.
