XGSwap: eXtreme Gradient boosting Swap for Routing in NISQ Devices

TL;DR

This paper introduces XGSwap, a machine learning-based routing method for NISQ devices that predicts higher fidelity paths, improving quantum state transfer despite noise and connectivity limitations.

Contribution

It presents a novel gradient boosting model to select better routing paths in quantum devices, outperforming traditional shortest-path methods.

Findings

Successfully identified higher fidelity paths in 23% of tests

Trained on 4050 CNOT gate samples ranging from 2 to over 100 qubits

Demonstrated effectiveness on a 127-qubit IBM quantum system

Abstract

In the current landscape of noisy intermediate-scale quantum (NISQ) computing, the inherent noise presents significant challenges to achieving high-fidelity long-range entanglement. Furthermore, this challenge is amplified by the limited connectivity of current superconducting devices, necessitating state permutations to establish long-distance entanglement. Traditionally, graph methods are used to satisfy the coupling constraints of a given architecture by routing states along the shortest undirected path between qubits. In this work, we introduce a gradient boosting machine learning model to predict the fidelity of alternative--potentially longer--routing paths to improve fidelity. This model was trained on 4050 random CNOT gates ranging in length from 2 to 100+ qubits. The experiments were all executed on ibm_quebec, a 127-qubit IBM Quantum System One. Through more than 200+ tests run on actual hardware, our model successfully identified higher fidelity paths in approximately 23% of cases.