Superconducting integrated circuits have emerged as a promising architecture for building quantum computers. However, a significant challenge is decoherence caused by spurious atomic tunneling defects at qubit electrode interfaces. These defects, acting as two-level systems (TLS), resonantly absorb energy from the qubit's oscillating electric field, reducing the energy relaxation time (T₁). Existing processors often utilize transmon qubits based on discrete energy levels in non-linear LC resonators. Dielectric loss in the capacitor electrodes, stemming from these TLS defects, is a major contributor to decoherence. The frequency dependence of this loss originates from individual resonances of these TLS, which can fluctuate due to interactions with other low-energy TLS. This fluctuation transforms thermal noise into the qubit's environmental spectrum, impacting the qubit's resonance frequency and relaxation rate, thereby affecting the quantum volume (computational power). Recent research demonstrated that the resonance frequencies of TLS on thin-film electrodes can be tuned by an applied DC-electric field. This allows tuning defects away from qubit resonance, leading to longer T₁ times. This paper presents a method to automatically find an optimal electric field bias that maximizes T₁ time, experimentally validating the concept and paving the way for in situ coherence optimization in quantum processors.

The paper draws upon existing literature highlighting the challenges of decoherence in superconducting qubits, particularly the role of two-level systems (TLS) in dielectric loss. It references previous work on the frequency dependence of dielectric loss and the structured nature of the loss originating from individual TLS resonances. The authors cite research on the impact of TLS interactions with thermally activated fluctuators, leading to fluctuations in qubit resonance frequency and relaxation rates. Previous findings demonstrating the ability to tune TLS resonance frequencies using an applied DC-electric field are also acknowledged, laying the foundation for the current study's approach.

The experiments used a transmon qubit sample in an 'X-Mon' design. A DC-electrode, positioned above the qubit chip, generated the electric field for TLS tuning. The response of TLS to the applied electric field was observed by measuring the qubit energy relaxation time (T₁) as a function of qubit frequency, revealing Lorentzian minima at TLS resonances. Swap-spectroscopy was used to obtain a detailed view of the TLS spectrum. The method for optimizing the qubit T₁ time involved measuring T₁ for a range of applied electric fields at various qubit resonance frequencies. The data was smoothed to identify broader maxima corresponding to more stable improvements. The E-field was then set to the value maximizing T₁, approaching it from the same starting value to mitigate hysteresis effects. A second, finer sweep was conducted around the optimum value to refine the results. To benchmark the method, T₁ was repeatedly measured at zero applied electric field and compared with measurements at the optimized E-field over a 30-minute interval. This benchmarking was repeated at various qubit resonance frequencies (59 in total). The average improvement of the T₁ time after optimization was calculated from the full dataset presented in the supplementary information.

The optimization routine consistently improved the 30-minute averaged qubit T₁ time in the majority (85%) of cases across various qubit resonance frequencies. A significant improvement (greater than 10%) was observed in 67% of cases, while in 46% of the tests, the T₁ time was enhanced by more than 20%. The average improvement over all tested frequencies and a 30-minute time interval was approximately 23%. The study also investigated the effect of the optimization on the temporal fluctuations of T₁. While some increase in fluctuation was observed in slightly more than half the cases (59%), this suggests a potential for algorithm improvement by prioritizing broader T₁ peaks. The optimization routine typically took less than 10 minutes, with potential for further reduction by optimizing data acquisition strategies. Analysis of raw data suggested that the optimization algorithm could be improved by including the width of T₁ peaks as a criterion, and by averaging over multiple E-field sweeps to reduce the impact of fluctuating TLS. A linear or machine learning feedback mechanism could also be implemented to continuously adjust the E-field bias based on qubit error rates.

The results demonstrate the effectiveness of tuning TLS resonances away from the qubit frequency using a DC-electric field to enhance qubit coherence. The significant average improvement in T₁ time (23%) underscores the potential of this method for improving the performance of superconducting qubits. The observed fluctuations in T₁ after optimization highlight the dynamic nature of the TLS interactions and suggest opportunities for algorithm refinement to improve stability. The relatively short optimization time (less than 10 minutes, potentially reducible to under one minute) makes this technique practically feasible for use in quantum processors. The method is particularly beneficial for fixed-frequency qubits, where tuning options are limited. Although the method improves T₁, further research is needed to quantify its impact on the dephasing time T₂.

This work presents a successful experimental demonstration of enhancing the coherence of superconducting transmon qubits by manipulating TLS defects using an applied DC-electric field. The automated optimization routine yielded an average T₁ time improvement of 23%, showcasing a promising technique for enhancing the performance of superconducting quantum processors. Future work could focus on refining the optimization algorithm to improve stability and reduce optimization time, as well as quantifying the impact on dephasing times. The integration of local gate electrodes for simultaneous in situ optimization of all qubits in a processor is a compelling direction for future research.

The study's findings were based on a specific qubit design and experimental setup. The observed improvements in T₁ might vary depending on factors like fabrication techniques, materials, and defect distributions. The 30-minute averaging period for T₁ measurement could have limited the ability to fully capture the dynamic behavior of TLS fluctuations. The method's effectiveness might also depend on the specific characteristics of the dominant TLS in each qubit.