15 May 2025

The AI system AlphaTensor-Quantum can make quantum computing more efficient. By optimizing quantum circuits, it makes calculations more efficient to save resources and accelerate discoveries.

The paper 'Quantum Circuit Optimization with AlphaTensor' was recently published in the journal Nature Machine Intelligence. Dr. John van de Wetering of the Theoretical Computer Science (TCS) group of the Informatics Institute of the UvA is involved in this research project together with researchers of Google DeepMind and quantum computing company Quantinuum. In the paper the authors describe a new AI-based tool called AlphaTensor-Quantum that was trained using reinforcement learning to make quantum computations more efficient. AlphaTensor-Quantum outperforms existing methods of quantum optimization, and matches the best human-designed solutions across multiple applications. This shows the potential of using machine learning for designing complicated quantum computations to help accelerate discoveries, saving on resources and time.

In 2022 DeepMind released their AlphaTensor paper where they showed an AI system can learn to improve matrix multiplication algorithms by factoring a complicated mathematical object called a tensor in smart ways.

Van de Wetering and the coauthors Konstantinos Meichanetzidis and Tuomas Laakkonen (Quantinuum) immediately realised that this technique should also be adaptable to optimise quantum circuits instead (specifically an important metric known as the T-count). We approached DeepMind with this proposal and they were immediately onboard. It was exciting to see that our hunch was correct, and that this method indeed can lead to significant improvements in performance for certain classes of quantum circuits, although it took a considerable intellectual and engineering effort by the teams from DeepMind and Quantinuum to get to this point.

We immediately realised that this technique should also be adaptable to optimise quantum circuits instead. We approached DeepMind with this proposal and they were immediately onboard. It was exciting to see that our hunch was correct

John van de Wetering, UvA