DOI

https://doi.org/10.1039/D3DD00047H

Abstract

Ising machines are hardware-assisted discrete optimizers that often outperform purely software-based optimization. They are implemented, e.g., with superconducting qubits, ASICs or GPUs. In this paper, we show how Ising machines can be leveraged to gain efficiency improvements in automatic molecule design. To this end, we construct a graph-based binary variational autoencoder to obtain discrete latent vectors, train a factorization machine as a surrogate model, and optimize it with an Ising machine. In comparison to Bayesian optimization in a continuous latent space, our method performed better in three benchmarking problems. Two types of Ising machines, a qubit-based D-Wave quantum annealer and GPU-based Fixstars Amplify, are compared and it is observed that the GPU-based one scales better and is more suitable for molecule generation. Our results show that GPU-based Ising machines have the potential to empower deep-learning-based materials design.