Details are available at our published paper https://doi.org/10.1039/D3DD00047H.
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 molecules design. To this aim, 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, qubit-based D-Wave quantum annealer and GPU-based Fixstars Amplify, are compared to observe that GPU-based one scales better and more suitable for molecule generation. Our results show that GPU-based Ising machines have the potential to empower deep-learning-based materials design.
The implementation of binary VAE is based on the work Junction Tree Variational Autoencoder for Molecular Graph Generation.
amplify==0.9.1
joblib==1.1.0
matplotlib==3.5.2
networkx==2.6.3
numexpr==2.8.1
numpy==1.21.5
rdkit==2022.9.5
scikit_learn==1.2.1
scipy==1.7.3
torch==1.11.0
tqdm==4.64.0
data/contains codes for randomly selecting labeled data. Please refer todata/README.mdfor details.train/contains codes for binary VAE training. Please refer totrain/README.mdfor details.bJTVAE/contains codes for the binary VAE model implementation.im/contains codes for optimizing latent binary molecular space via an Ising machine. Please refer toim/README.mdfor details.
Zetian Mao ([email protected])
Department of Computational Biology and Medical Science
The University of Tokyo
Cite this code:
