TY - JOUR
T1 - Hybrid Quantum Neural Network Model with Catalyst Experimental Validation
T2 - Application for the Dry Reforming of Methane
AU - Roh, Jiwon
AU - Oh, Seunghyeon
AU - Lee, Donggyun
AU - Joo, Chonghyo
AU - Park, Jinwoo
AU - Moon, Il
AU - Ro, Insoo
AU - Kim, Junghwan
N1 - Publisher Copyright:
© 2024 American Chemical Society.
PY - 2024/3/11
Y1 - 2024/3/11
N2 - Machine learning (ML), which has been increasingly applied to complex problems such as catalyst development, encounters challenges in data collection and structuring. Quantum neural networks (QNNs) outperform classical ML models, such as artificial neural networks (ANNs), in prediction accuracy, even with limited data. However, QNNs have limited available qubits. To address this issue, we introduce a hybrid QNN model, combining a parametrized quantum circuit with an ANN structure. We used the catalyst data sets of the dry reforming of methane reaction from the literature and in-house experimental results to compare the hybrid QNN and the ANN models. The hybrid QNN exhibited superior prediction accuracy and a faster convergence rate, achieving an R2 of 0.942 at 2478 epochs, whereas the ANN achieved an R2 of 0.935 at 3175 epochs. For the 224 in-house experimental data points previously unreported in the literature, the hybrid QNN exhibited an enhanced generalization performance. It showed a mean absolute error (MAE) of 13.42, compared with an MAE of 27.40 for the ANN under similar training conditions. This study highlights the potential of the hybrid QNN as a powerful tool for solving complex problems in catalysis and chemistry, demonstrating its advantages over classical ML models.
AB - Machine learning (ML), which has been increasingly applied to complex problems such as catalyst development, encounters challenges in data collection and structuring. Quantum neural networks (QNNs) outperform classical ML models, such as artificial neural networks (ANNs), in prediction accuracy, even with limited data. However, QNNs have limited available qubits. To address this issue, we introduce a hybrid QNN model, combining a parametrized quantum circuit with an ANN structure. We used the catalyst data sets of the dry reforming of methane reaction from the literature and in-house experimental results to compare the hybrid QNN and the ANN models. The hybrid QNN exhibited superior prediction accuracy and a faster convergence rate, achieving an R2 of 0.942 at 2478 epochs, whereas the ANN achieved an R2 of 0.935 at 3175 epochs. For the 224 in-house experimental data points previously unreported in the literature, the hybrid QNN exhibited an enhanced generalization performance. It showed a mean absolute error (MAE) of 13.42, compared with an MAE of 27.40 for the ANN under similar training conditions. This study highlights the potential of the hybrid QNN as a powerful tool for solving complex problems in catalysis and chemistry, demonstrating its advantages over classical ML models.
KW - artificial neural network
KW - catalyst
KW - data-driven modeling
KW - dry reforming of methane
KW - machine learning
KW - parameterized quantum circuit
KW - quantum neural network
UR - http://www.scopus.com/inward/record.url?scp=85186208184&partnerID=8YFLogxK
U2 - 10.1021/acssuschemeng.3c07496
DO - 10.1021/acssuschemeng.3c07496
M3 - Article
AN - SCOPUS:85186208184
SN - 2168-0485
VL - 12
SP - 4121
EP - 4131
JO - ACS Sustainable Chemistry and Engineering
JF - ACS Sustainable Chemistry and Engineering
IS - 10
ER -