GPU-accelerated approximate kernel method for quantum machine learning.

Area of Science:

• Computational chemistry
• Quantum machine learning
• Materials science

Background:

• Accurate prediction of molecular energies and forces is crucial for atomistic simulations.
• Traditional methods can be computationally expensive, limiting simulation timescales.
• Quantum Machine Learning (QML) offers a promising avenue for accelerating these predictions.

Purpose of the Study:

• To introduce QML-Lightning, a PyTorch package for GPU-accelerated QML.
• To enable rapid training of approximate kernel models for energy and force predictions.
• To demonstrate the efficiency and accuracy of QML-Lightning on benchmark datasets.

Main Methods:

• Development of QML-Lightning, a PyTorch package leveraging GPU acceleration.
• Implementation of a cost-efficient GPU version of the FCHL19 kernel model.
• Benchmarking using QM9, MD-17, and 3BPA datasets for atomistic simulations.

Main Results:

• QML-Lightning achieves trained models within seconds.
• The package provides energy and force predictions with competitive accuracy.
• Microsecond per atom timescale simulations are feasible with QML-Lightning.
• Learning curves and timings are reported for legacy benchmarks.

Conclusions:

• QML-Lightning significantly accelerates QML model training and inference.
• The package offers a practical solution for high-throughput atomistic simulations.
• GPU-accelerated approximate kernel models are effective for materials science applications.