A reinforcement learning network can now autonomously discover high-performance pulse sequences using solely system parameters and a reward function. This approach applies generally to superconducting qubit platforms, requiring no ansatz and operating at nanosecond resolution consistent with hardware limitations. Applying this method to ground-state preparation of the Hydrogen molecule on a simulated superconducting device serves as proof of concept. The agent reliably finds optimised control sequences achieving high fidelity and exceeding the performance of random search methods. These findings indicate that adaptive learning is a strong, hardware-ready framework for pulse-level quantum control.

Autonomous reinforcement learning optimises qubit control exceeding previous fidelity benchmarks

Error rates in preparing the Hydrogen molecule’s ground state dropped to below 1×10−3 using this new method, representing an order of magnitude improvement compared to conventional random search techniques which previously struggled to consistently achieve such low error levels. Achieving these fidelities is key for scaling up molecular calculations on near term devices; it enables quantum simulations demanding high precision that were unattainable with prior pulse optimisation strategies.

Southern Methodist University researchers successfully demonstrated a reinforcement learning system capable of autonomously designing microwave pulses, the signals used to manipulate qubits, without relying on pre-defined circuit layouts or detailed models of the superconducting device itself.

A deep Q learning network operates at nanosecond resolution, aligning well with current hardware capabilities and enabling rapid pulse shaping. It requires no initial ‘guess’ about what good control sequences might look like, but these results currently rely on simulations rather than physical hardware implementation, scaling up to larger molecules will demand further improvements in computational efficiency and strong durability against real-world noise sources.

Automated pulse design overcomes calibration challenges in advanced qubit manipulation

The team has developed an innovative approach to controlling qubits, the fundamental building blocks of quantum computers, by automating microwave pulse design with exceptional precision; this tackles the growing challenge of calibrating increasingly complex systems exhibiting unpredictable behaviour and imperfections. While current methods, such as ctrl-VQE, attempt direct pulse optimisation, they often become trapped searching vast control fields. Efficient strategies are required for navigating these complexities. Fully automating quantum control remains a formidable challenge given inherent noise and system complexity but acknowledging this is important; this work represents major progress towards practical implementation nonetheless.

This automated approach consistently achieves high fidelity and outperforms conventional methods like random search. The team’s successful demonstration marks a shift from manually calibrated quantum control toward adaptable machine learning techniques, simplifying optimisation across diverse superconducting qubit platforms by bypassing the need for pre-defined circuit layouts known as ansätze. Achieving high fidelity state preparation without relying on detailed system models is particularly valuable considering increasing hardware complexity and imperfections within devices operating at nanosecond timescales; it offers significant advantages in managing these challenges effectively.

The researchers successfully trained a reinforcement-learning agent to design microwave pulses that prepare quantum states with high accuracy. This method autonomously discovers effective pulse sequences using only device parameters and a reward function, outperforming random search approaches when applied to simulations of hydrogen molecule ground-state preparation. Because this technique requires no initial assumptions about optimal control strategies or predefined circuits, it presents an adaptable framework for controlling superconducting qubits. The authors note further computational efficiency improvements will be needed to scale the approach and address real-world noise sources.

👉 More information
🗞 Adaptive AI for Pulse-Level Quantum Control
✍️ Sanjeev Shapkota, Yayu Mo and Sanjaya Lohani
🧠 ArXiv: https://arxiv.org/abs/2609.08727

More like this