Unleashing the Potential of LLMs for Quantum Computing: A Study in Quantum Architecture Design

TL;DR

Addressing the challenge of designing quantum architectures for $VQAs$ in the $NISQ$ era and into $FTQC$, this work presents $QGAS$, a GPT-guided architecture search that iteratively proposes, trains, and evaluates ansatz circuits. The framework combines $GPT$-4 driven design with $QASM$ output from $GPT$-3.5, maps problem Hamiltonians (via fermionic mapping for chemistry or Ising mapping for finance), and optimizes parameters with gradient-based methods, all under a human-in-the-loop feedback mechanism. Across benchmarks in quantum chemistry (e.g., $H_2O$, $LiH$) and quantum finance (portfolio optimization, MaxCut, TSP), $QGAS$ achieves high-performance ansätze with limited prompts, often matching or surpassing state-of-the-art baselines like $QuantumNAS$ in noiseless settings and showing improved efficiency when guided by human input. The study also highlights limitations of current $GPT$ models—such as noise sensitivity and potential biases—and outlines practical implications and future directions for integrating $GPT$ into quantum-architecture research and fault-tolerant design.

Abstract

Large Language Models (LLMs) contribute significantly to the development of conversational AI and has great potentials to assist the scientific research in various areas. This paper attempts to address the following questions: What opportunities do the current generation of generative pre-trained transformers (GPTs) offer for the developments of noisy intermediate-scale quantum (NISQ) technologies? Additionally, what potentials does the forthcoming generation of GPTs possess to push the frontier of research in fault-tolerant quantum computing (FTQC)? In this paper, we implement a QGAS model, which can rapidly propose promising ansatz architectures and evaluate them with application benchmarks including quantum chemistry and quantum finance tasks. Our results demonstrate that after a limited number of prompt guidelines and iterations, we can obtain a high-performance ansatz which is able to produce comparable results that are achieved by state-of-the-art quantum architecture search methods. This study provides a simple overview of GPT's capabilities in supporting quantum computing research while highlighting the limitations of the current GPT at the same time. Additionally, we discuss futuristic applications for LLM in quantum research.

Figures (5)

• Figure 1: Illustration of design and implementation of QGAS. Following an initial encoding issue, GPT-4 suggests an ansatz, which is subsequently processed by a sub-model constructed by GPT-3.5 to convert it into QASM format. To assess the effectiveness of the ansatz, a benchmarking application is executed, enabling the evaluation of its quality. The obtained results are then fed back to GPT-4 through a natural language prompt, facilitating further iterations and refinements.
• Figure 2: Visualization of benchmark applications. a) Max-Cut problem with 5 nodes. b) Traveling Salesman Problem with 3 nodes. c) Portfolio Optimization problem. d) Molecule structure of $H_{2}O$. e) Molecule Structure of $LiH$.
• Figure 3: Experiments for two trials of the portfolio optimization problem and evaluation of the ansatz architecture generated by QGAS. We show both the gatecounts and estimated value for each iteration, a lower estimated value with smaller gatecounts indicates a better ansatz. In both trials, the results demonstrate in a limited number of prompt guidelines and iterations, we can obtain a high-performance ansatz generated by QGAS.
• Figure 4: Experiments for two trials of the quantum chemistry molecule ground state energy tasks and evaluation of the ansatz architecture generated by QGAS. We show both the epochs and estimated energy for each iteration, where a lower estimated energy with fewer epochs indicates a better ansatz.
• Figure 5: The application benchmarks for comparing the state-of-art ansatz and QGAS-generated ansatz. a) Molecule ground state energy estimation tasks for $H_2O$ and $LiH$, compare the ansatz generated by QGAS, UCCSD, and QuantumNAS. b) Machine learning tasks for MNIST-2 and MNIST-4 classification, compare the QGAS-generated ansatz, random generated ansatz, and QuantumNAS.