Accelerating Parameter Initialization in Quantum Chemical Simulations via LSTM-FC-VQE

TL;DR

This paper introduces an LSTM-based meta-learning framework to improve parameter initialization in quantum chemical simulations, significantly reducing convergence time and energy errors in VQE methods for larger molecules.

Contribution

It proposes a novel LSTM-FC-VQE architecture that learns from small molecules to efficiently initialize parameters for larger systems, enhancing quantum simulation performance.

Findings

Faster convergence in VQE simulations.

Lower energy errors compared to traditional methods.

Effective across molecules of varying sizes.

Abstract

We present a meta-learning framework that leverages Long Short-Term Memory (LSTM) neural networks to accelerate parameter initialization in quantum chemical simulations using the Variational Quantum Eigensolver (VQE). By training the LSTM on optimized parameters from small molecules, the model learns to predict high-quality initializations for larger systems, reducing the number of required VQE iterations. Our enhanced LSTM-FC-VQE architecture introduces a fully connected layer, improving adaptability across molecules with varying parameter sizes. Experimental results show that our approach achieves faster convergence and lower energy errors than traditional initialization, demonstrating its practical potential for efficient quantum simulations in the NISQ era.