Optimizing restricted Boltzmann machine learning by injecting Gaussian noise to likelihood gradient approximation

Sanjaya, Prima; Kang, Dae-Ki

doi:10.1007/s10489-018-01400-5

Cited by 1 publication

(2 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The hidden layer converts the input information into hidden vectors through the activation function and transfers them to the output layer, and the output layer outputs the predicted H 2 S solubility value. Mathematically, for the given input vector X , the output of the two hidden layers and output layers of DNN is calculated as follows F 1 ( a 1 ) = F 1 ( W 1 × X + b 1 ) F 2 ( a 2 ) = F 2 ( W 2 × F 1 false( a 1 false) + b 2 ) F 3 ( a 3 ) = F 3 ( W 3 × F 2 false( a 2 false) + b 3 ) where, F 1 , F 2 , and F 3 are two hidden layer and output layer activation functions, respectively. b 1 , b 2 , and b 3 are bias terms for neurons in the hidden and output layers, respectively.…”

Section: Modelingmentioning

confidence: 99%

Section: ■ Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Prediction of the Solubility of Acid Gas Hydrogen Sulfide in Green Solvent Ionic Liquids via Quantitative Structure–Property Relationship Models Based on the Molecular Structure

Liu

Dong

Zhu

et al. 2023

ACS Sustainable Chem. Eng.

View full text Add to dashboard Cite

Ionic liquids (ILs) can be used as capturing acidic gases that damage the environment. By establishing a quantitative structure–property relationship (QSPR) model of the IL structure, temperature, pressure, and H2S solubility, it can be used to screen ILs with excellent properties. In this study, molecular descriptors (MD), molecular identifiers (MI), and combinations of MD and MI (MD_MI) are used to represent the structure of ILs, combining pressure and temperature as the input of the model; the QSPR model is built by coupling the two models of the deep neural network (DNN) and random forest (RF). We find that the model constructed by using MI to represent ILs and DNN has the best performance. The Shapley additive explanation (SHAP) method is used to analyze features to obtain the most valuable molecular structure information for prediction. The results show that the contribution degree and direction of different MD and MI in the prediction of H2S solubility are different and correctly identify the impact of environmental factors (temperature and pressure). Finally, the electrostatic potential (ESP) between H2S and ILs was calculated to study the influence of different carbon chain lengths on the solubility of H2S.

show abstract