In quantum machine learning, circuits on a quantum computer serve as learning models that are trained to represent relationships in data. Similar to classical machine learning, their parameters are adjusted incrementally to improve the results. This is typically done using the parameters of predefined operations on a quantum computer, known as gates. To put it simply, the training process is like searching for the lowest point in a landscape of mountains and valleys, where the optimization process can become trapped in valleys even though better solutions exist elsewhere. Such local minima thus complicate the training process and prevent the model’s full potential from being realized.
Researchers at the Karlsruhe Institute of Technology (KIT) and OTH Regensburg are therefore investigating additional degrees of freedom that arise from the physical implementation of the gates on the quantum computer. During execution, gates are translated into microwave pulses, and their properties, such as amplitude or duration, can be incorporated into the training process. When applied to the landscape described above, this opens additional directions of movement that allow the training to escape a valley. This approach is particularly relevant for gates composed of several simpler operations, which therefore involve coupled parameters. If such parameters are adjusted independently of one another, the model gains more freedom to adapt its output to the training data.
The research team was able to describe this mechanism mathematically and confirm it in numerical experiments using quantum Fourier models. Such models represent relationships as a superposition of oscillations of different frequencies much like a sound is composed of multiple tones. In the simulations, models were tasked with learning given curves; incorporating pulse parameters led, on average, to smaller errors across all circuit variants examined. The additional parameters alter how the available frequency components can be utilized and weighted. Which frequencies are in principle available is determined by the encoding of the data.
The results thus demonstrate how the physical control of a quantum computer can actively contribute to the learning process. Since previous studies are based on simulations that do not include noise and other imperfections present in real quantum devices, the next step is to evaluate the concept on quantum hardware. This will reveal how well the improved training capabilities can be reconciled with the technical limitations of the pulses and the additional effort required for their optimization.
For their paper “Beyond Gates: Pulse Level Quantum Fourier Models” (preliminary arXiv DOI), the team was awarded third place in the Quantum Machine Learning division for the Best Paper Award at this year’s IEEE Quantum Week conference, among 50 accepted papers.
The IEEE International Conference on Quantum Computing and Engineering (QCE), also known as IEEE Quantum Week, is an annual international conference covering quantum computing from scientific fundamentals to industrial applications. In 2026, it brought together over 2,300 participants for talks, workshops, and an accompanying exhibition.
Contact at SCC: Melvin Strobl
Caption: Melvin Strobl is delighted to receive a Best Paper Award. Pictured at the award ceremony (from left to right): Candace Culhane, Los Alamos National Laboratory, QCE25 General Chair; Arno Jacobsen, University of Toronto, QCE26 QML Technical Papers Track Chair; Melvin Strobl, KIT/SCC; Grace Lewis, Software Engineering Institute, Carnegie Mellon University, 2026 President of the IEEE Computer Society.