Component-Level Inverse Design of Transmon Qubits Using Neural Networks

Designing superconducting qubits typically requires a slow, iterative forward-design loop where engineers manually adjust layout geometries, run time-intensive electromagnetic (EM) simulations, and compute Hamiltonian parameters until the desired qubit properties are met. This paper addresses the need for a more efficient, systematic approach by developing an inverse-design workflow that directly maps target Hamiltonian parameters (qubit frequency and anharmonicity) to the physical layout parameters required to realize them.

The researchers employed a tandem neural-network architecture. First, they trained a forward surrogate model—a multilayer perceptron (MLP)—to accurately predict Hamiltonian parameters from geometry inputs. Once trained, this surrogate was frozen and placed after an inverse MLP. During training, the inverse model learns to map target Hamiltonian values to geometry parameters by minimizing the difference between the target Hamiltonian and the Hamiltonian reconstructed by the frozen surrogate. This approach avoids the common pitfalls of direct geometry-regression, where multiple valid physical designs can lead to poor, averaged results.

The inverse-design pipeline proved highly efficient and accurate. In validation against conventional EM solvers, 97% of the generated designs were physically usable. The model achieved mean percent errors of 0.73% for qubit frequency and 1.58% for anharmonicity, which are comparable to or better than the uncertainties inherent in standard academic fabrication and simulation processes. Furthermore, the pipeline demonstrated massive computational speedups: a single query takes approximately 56 ms on a CPU, compared to roughly 2 minutes for a conventional EM solver. When using batching, the runtime per sample drops to 2.5 microseconds on a GPU, representing a speedup of nearly 48 million times over a single conventional CPU-based EM extraction.