Scientists are studying variational quantum circuits as a promising machine learning model, but achieving optimal performance requires careful circuit design, which is often a difficult and time-consuming process. Grier M. Jones and Viki Kumar Prasad, along with Edward S. Rogers Sr. School of Electrical and Computer Engineering, University of Toronto, Canada, and School of Chemical and Physical Sciences, University of Toronto Mississauga, Canada, and Aviraj Newatia, Edward S. Rogers Sr. School of Electrical and Computer Engineering, University of Toronto, Canada, School of Computer Science, University of Toronto, Canada, and Vector Institute for Artificial Institute. Intelligence, Canada, and colleagues have presented a new evolution-inspired algorithm to optimize these circuits through local gate changes. This research, conducted in collaboration with researchers in the Department of Chemistry at the University of Calgary, Canada, introduces a method to automatically discover competing circuit architectures, which has been demonstrated through successful application to synthetic regression tasks and complex datasets such as bond separation energies and water conformer data. The ability to efficiently design high-performance quantum circuits is an important step toward realizing the potential of quantum machine learning and deploying these models on today’s hardware.
Parameterized quantum circuits are flexible, but often require painstaking manual design to achieve optimal performance for a particular task. By applying a fixed set of gate-level actions to existing circuits, the algorithm efficiently explores promising configurations.
This local search strategy is motivated by the observation that many effective quantum circuits can be derived from relatively small perturbations of already working designs. This performance metric is computed through state vector simulation and indicates the frequency of incorrect predictions made by the model during each computational step.
Analysis of the discovered circuits reveals that the algorithm prioritizes structure preservation during refinement, allowing targeted improvements while maintaining functional integrity. The best-performing model was successfully deployed on state-of-the-art quantum hardware, validating its practical applicability beyond simulation. This development confirms the feasibility of converting algorithmically designed circuits into concrete quantum computations.
This approach avoids the limitations of previous quantum architecture search methods, which often suffered from the computational cost of searching vast configuration spaces. Additionally, a data set of water conformers generated using a data-driven coupled cluster approach provided a challenging benchmark to evaluate the capabilities of the algorithm in modeling molecular properties.
The choice of this dataset reflects the potential of quantum machine learning to accelerate computationally intensive tasks in chemistry and materials science. However, algorithms alone are not sufficient to realize this potential. Efficient design of quantum circuits for specific tasks is required.
This work represents an important step toward automating that process and demonstrates how quantum circuit architectures can be evolved through local stochastic exploration. Furthermore, the fixed set of gate-level actions can limit the exploration of truly new circuit topologies. In the future, it is expected that these architectural search algorithms and techniques for optimizing circuit compilation and error correction will converge. The next generation of tools will not only find better circuits, but also build them, adapting them to the specific constraints of available hardware and pushing the boundaries of what is computable.
