Automatically Discovering and Optimizing Quantum Architectures

With Quantum Architecture Search (QAS), we are developing a method that approximates user-defined functions using parameterized quantum circuits. Our framework learns circuit structures and input encodings directly from example data, taking into account both accuracy and the required computational resources. For many tasks, there is no single ideal quantum circuit. Accuracy, computational effort, and hardware requirements vary depending on the application. We are developing methods that automatically find and optimize suitable quantum architectures. This enables new quantum applications to be developed more quickly and better adapted to available resources.

To execute a function on a quantum computer, a suitable circuit structure and an appropriate input encoding are required. The circuit should operate as precisely as possible while keeping the required resources – such as circuit depth or the number of two-qubit gates – to a minimum. Depending on the function and the input range, a different architecture may be optimal.

With QAS, we automatically search for suitable solutions based on representative input-output examples. Users do not need to specify a concrete quantum implementation. The achievable accuracy and the number of resources required depend on the specific task and the input range under consideration.

How Our Quantum Architecture Search Works

Our search begins with a population of simple candidates. An evolutionary algorithm gradually modifies the circuit structure and the arrangement of the input encodings. An classical optimizer then trains the variable parameters of the circuits.

For evaluation, we consider both the training error and the required circuit resources. We collect the best solutions in what is known as a Pareto front. It contains candidates that each represent different trade-offs between accuracy and resource consumption.

To describe the candidates, we use graphical representations such as quantum circuits, measurement patterns, and ZX diagrams. ZX diagrams are a graphical representation of quantum algorithms that reveal structural relationships. By making targeted changes to the underlying graphs, the algorithm explores new architectures. From these, we then extract circuits for training and resource evaluation.

What Sets Our Approach Apart: Joint Optimization of Architecture and Coding

Our method adapts both the type of data input and the underlying computational structure. This allows us to tailor the quantum circuit specifically to the respective application.

In making our selection, we simultaneously take into account approximation errors, circuit depth, and the number of gates. This results in various solution options that clearly illustrate the relationship between accuracy and resource requirements.