Artificial intelligence (AI) techniques, particularly large language models (LLMs), are revolutionizing various scientific fields, and quantum computing is no exception. According to the NVIDIA Technical Blog, a collaboration between NVIDIA, the University of Toronto, and Saint Jude Children’s Research Hospital is leveraging generative pre-trained transformers (GPTs) to design new quantum algorithms, including the Generative Quantum Eigensolver (GQE) technique.

Innovative GQE Technique

The GQE technique represents a significant advancement in the field of AI for Quantum. Developed using the NVIDIA CUDA-Q platform, GQE is the first method allowing users to employ their own GPT models to create complex quantum circuits. The CUDA-Q platform, with its emphasis on accelerated quantum supercomputing, provides a hybrid computing environment ideally suited for the GQE technique.

Alan Aspuru-Guzik, a co-author of the GQE method, highlights the scalable nature of the CUDA-Q platform, which integrates CPUs, GPUs, and QPUs for training and using GPT models in quantum computing.

Understanding GQE through LLM Analogy

Conventional LLMs, which understand and generate text by learning from large datasets, provide a useful analogy for comprehending GQE. While LLMs deal with words, GQE deals with quantum circuit operations. GQE uses a transformer model to generate sequences of unitary operations, which define quantum circuits. The training involves minimizing a cost function evaluated by computing expectation values of previously generated circuits.

Advancements in NISQ Era Algorithms

In the noisy, small-scale quantum (NISQ) era, quantum algorithms face several hardware constraints. Hybrid quantum-classical algorithms like the Variational Quantum Eigensolver (VQE) aim to overcome these limitations. The GQE method, however, is the first to leverage AI to accelerate NISQ applications, optimizing parameters classically within the GPT model.

GQE extends NISQ algorithms by:

• Building quantum circuits without quantum variational parameters.
• Improving quantum resource efficiency by replacing quantum gradient evaluation with sampling and backpropagation.
• Allowing customization to incorporate domain knowledge or target applications outside of chemistry.
• Enabling pretraining to eliminate the need for additional quantum circuit evaluations.

Results and Future Prospects

For its inaugural application, the GQE model, inspired by GPT-2 and referred to as GPT-QE, was used to estimate ground state energies of small molecules. The model demonstrated significant parallelization capabilities, reducing training times from 173 hours on a single NVIDIA H100 GPU to 3.5 hours across 48 H100 GPUs.

Future research aims to explore different operator pools for GQE and optimal training strategies, including pretraining. This process uses existing datasets to enhance the efficiency of transformer training or aid in its convergence.

Beyond quantum chemistry, NVIDIA and Los Alamos National Lab are investigating the application of GQE to geometric quantum machine learning, showcasing the broad potential of this innovative approach.

For more information about this collaboration and the GQE code, visit the NVIDIA Technical Blog.

Image source: Shutterstock