Evrim Yazgin
Cosmos science journalist
Engineers at Australia’s national science agency, CSIRO, have performed a world-first use of quantum machine learning to fabricate semiconductors. The research could reshape the way future microchips are designed.
Head of quantum systems research at CSIRO’s Data61, Professor Muhammad Usman, has previously explained to Cosmos that quantum algorithm research today is critical for the development of useful quantum computers of tomorrow.
Usman explained that quantum machine learning (QML) has the potential to outperform classical machine learning (CML) algorithms.
This has been put to the test in the new research which is published in the journal Advanced Science. The study is the first to show that a quantum method can be applied to real experimental data in semiconductor fabrication to improve the process.
The team was particularly interested in modelling the Ohmic contact resistance of the semiconductor material. This property is a measure of the electrical resistance where the semiconductor comes into contact with a metal and the current flows easily between the materials in both directions.
Modelling Ohmic contact resistance is critical to semiconductor design and fabrication, but it’s also a property which is notoriously difficult to model.
They tested their QML model on data from 159 experimental samples of GaN HEMT (gallium nitride high electron mobility transistor) semiconductors. GaN HEMT offers superior performance compared to the more common silicon-based transistors.
“Once we get the semiconductor fabrication data sets, we do a lot of pre-processing. This pre-processing is classical step. We take different parameters which influence fabrication and do a sort of ‘hot encoding’ which basically just says whether a particular parameter is triggered or not triggered,” explains Usman who is the senior author of the new study.
“It’s 1s and 0s which just indicate whether, for example, the particular gas was turned on or not on, annealing time, whether it was doped or not,” he says.
Once the hot encoding was done, the team had a list of 37 parameters for each experiment. A further classical analysis, called principal component analysis, reduced the parameters to just 5.
“The quantum computers that we currently have are very limited capabilities. So we wanted to simplify it. We wanted to make sure that we can reduce the dimensionality of the problem intelligently, so that we can actually do it within the capabilities of the current quantum processes,” Usman says.
“Once we have done that, then we start the quantum component.”
The team developed an innovative Quantum Kernel-Aligned Regressor (QKAR) architecture.
Their QKAR setup included a Pauli-Z quantum feature map – a mathematical operator which can translate classical data into quantum states in the form of 5 quantum bits, or qubits.
Once data is mapped to the qubits, a quantum kernel alignment layer is used to perform the machine learning.
In computing, kernels are the core components of the operating system. They manage the system’s resources and bridge between the software and hardware elements.
Usman explains that the quantum kernel calculations extract the important features from the fabrication data sets.
“That is where all the quantum magic is happening, because these kernels are highly entangled. When they process the data set, they access information that would not be otherwise available from the classical kernels that people have used in the past.”
After the quantum kernel has extracted the important features, a final classical algorithm is used to retrieve the information.
“This classical machine learning technique takes that outcome that the quantum method has extracted, and then it’s trained to guidance back to the fabrication. It can tell us what the important parameters in the fabrication process are which play the critical role and what needs to be changed or tuned to optimise fabrication,” Usman explains.
The QKAR technique outperformed 7 CML algorithms also trained on the same problem.
He adds that, because only 5 qubits are needed, the method is immediately applicable to current quantum architectures.
“So this is a very friendly technique. Normally when people talk about quantum algorithms, they require 10s of qubits which are not available. But this method that we have developed by combining classical and quantum can immediately, or in the near future, be implemented and get benefits.”
“The semiconductor industry is increasingly constrained by data scarcity and rising process complexity,” says lead author Dr Zeheng Wang.
“Our results show that quantum models, when carefully designed, can capture patterns that classical models may miss, especially in high-dimensional, small-data regimes. We validated the model by fabricating new GaN devices, which showed optimised performance, and, through quantum kernel spectrum analysis, confirming QML’s ability to generalise beyond training data.”
“One of the biggest challenges in quantum machine learning is making it practical,” says co-author Dr Tim van der Laan. “By introducing a learnable kernel alignment layer into a shallow quantum circuit, we’ve demonstrated that useful performance gains are achievable even with limited qubit hardware.”
“The model also showed robustness under realistic levels of quantum noise, which is essential for future implementation on actual NISQ (Noisy intermediate-scale quantum) devices.”
Usman says the QKAR model can be adapted for other materials beyond this initial proof-of-concept test on GaN.
“It is an example where quantum is basically clearly showing that that it can extract features which are not otherwise available from classical,” he says. “This is the very first study that we have published, and we have demonstrated that it works. Now we are going to be working with other material development scientists and looking at new material systems. We will also be looking at other semiconductor materials, such as silicon fabrication processes.
“That’s our next step: look at other data sets, see what the extent of the applicability is of this method and verify it for a range of different experimental samples.”