Standard molar entropy estimation of binary composite oxides: A novel hybrid approach based on deep learning and hyperparameters optimization

The standard molar entropy (SME), a typical physical and chemical property of binary composite oxides (BCOs), is of great significance in materials science. Estimating the SME is however challenging, being unstable and difficult to precisely determine both experimentally and theoretically due to the complex internal structure and diverse interactions within BCOs. In this work, the data-driven approach was used to estimate the standard molar entropy of BCOs. Specifically, the crested porcupine optimizer (CPO) was introduced to fine-tune the hyperparameters of the bidirectional time convolutional network (BiTCN) and bidirectional gated recurrent unit (BiGRU). Results indicate that the performance of CPO-BiTCN-BiGRU-SME improves with increasing the training dataset proportion. When the training dataset proportion reaches 95 %, the root-mean-square error (RMSE), mean-absolute-percentage error (MAPE) and coefficient of determination (R 2 ) between the estimated and the experimental SME values are 15.49 J∙mol −1 ∙K −1 , 11.31 % and 0.89, respectively. Compared to CPO-BiTCN and long short-term memory (LSTM), CPO-BiTCN-BiGRU-SME achieves RMSE reductions of 10.19 J∙mol −1 ∙K −1 and 16.55 J∙mol −1 ∙K −1 , respectively, while the MAPE is reduced by 4.34 % and 8.31 %, respectively. Notably, CPO-BiTCN-BiGRU-SME outperforms various existing hybrid and single models. The superiority lies in the fact that CPO enables the novel hybrid approach to better adapt to SME data and enhance its generalization ability, BiTCN extracts local features and captures short-term dependencies efficiently, BiGRU handles long-term dependencies and integrates nature effectively. • Pioneered merging DL and hyperparameter optimization for estimating SME of BCOs. • Introduced CPO to optimize hyperparameters of BiTCN-BiGRU-SME for the first time. • Performance of CPO-BiTCN-BiGRU-SME rises notably with more training dataset of SME. • Compared with traditional methods, CPO-BiTCN-BiGRU-SME reduces RMSE and MAPE. • Offered a novel hybrid approach for facilitating materials discovery and design.