AI-powered Technique Could Help Quantum Computers Scale

Circuit optimization is a key challenge in developing future fault-tolerant quantum computers at scale. One type of quantum gate, known as a T gate, is a barrier to achieving this as they are computationally expensive. AI could hold the solution to minimizing how many T gates are used to implement a given quantum circuit.

Researchers from quantum computing firm Quantinuum teamed up with counterparts at Google DeepMind to see if an AI model could help reduce T gate counts at scale.

They created the AlphaTensor-Quantum algorithm, which uses deep reinforcement learning to optimize optimizing quantum circuits. It employs ‘gadgets,’ which are specific configurations that reduce the overall T-count.

The algorithm essentially finds the best way to arrange the operations of a quantum computer, which is how it performs operations, so it can solve complex problems using fewer steps. The AI writes the code that the quantum computer then executes – in effect, making quantum computing both faster and more effective.

The algorithm can also incorporate domain-specific knowledge into its optimization process, which enables it to outperform existing T gate reduction methods.

The researchers used Google DeepMind’s AlphaTensor model for creating new matrix multiplication algorithms – think AI that creates algorithms. Their efforts, outlined in a paper published last week, matched the best human-designed solutions.

Related:AI Meets Quantum: Leap Forward for AI Models to ‘Understand’

The jointly developed process can be applied across virtually all quantum computing platforms, with Quantinuum posting on the company’s LinkedIn page that the algorithm “has the potential to replace manual or hybrid manual-machine collaboration in automatically finding the best constructions (in terms of T-count).”

The publication of the paper with Google DeepMind follows the release of Quantinuum’s framework to improve how AIs learn. The company sought to extract learnings from quantum development to improve the ability to teach machines to understand concepts and images.