A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Improving Prediction Performance on Solute Parameters Using Multitask Relational Graph Convolutional Networks with Explicit Hydrogens
https://orcid.org/0009-0002-2696-5727
https://orcid.org/0009-0007-3724-5182
https://orcid.org/0000-0002-5756-3336
This study proposed multitask (MT) learning coupled with relational graph convolutional networks with attention weights (RGAN) and explicit hydrogens (abbreviated as MT-RGAN-H architecture) to construct prediction models on the solute parameters including excess molar refraction, dipolarity/polarizability, H-bond acidity (A), H-bond basicity (B), and logarithmic hexadecane–air partition coefficient. The resulting MT-RGAN-H model was proved to outperform single-task (ST) machine learning models, ST-RGAN models, MT-RGAN model, and previous models on the solute parameters. Importantly, the MT-RGAN-H model improved the prediction accuracy for A (5.6% increase) and B (4.3% increase) over the previous models. The MT-RGAN-H architecture also enables learning from multiple interrelated end points, even from end points with limited labeled cases, overcoming challenges of data insufficiency in some tasks. By visualizing the attention weights, the MT-RGAN-H model was well interpreted. The predicted solute parameters were further employed to predict six physicochemical parameters of chemicals, achieving better prediction accuracy over the previous optimal models. Therefore, this study provides an integrative “end-to-end” prediction scheme for the solute parameters, laying a foundation for accurately predicting the environmental partition behavior of chemicals. The MT-RGAN-H architecture can be further applied to construct prediction models on other interrelated end points, supporting sound management of chemicals.
A multitask relational graph convolutional network model with explicit hydrogen atoms and attention weights was developed for accurately predicting solute parameters of chemicals.
Improving Prediction Performance on Solute Parameters Using Multitask Relational Graph Convolutional Networks with Explicit Hydrogens
https://orcid.org/0009-0002-2696-5727
https://orcid.org/0009-0007-3724-5182
https://orcid.org/0000-0002-5756-3336
This study proposed multitask (MT) learning coupled with relational graph convolutional networks with attention weights (RGAN) and explicit hydrogens (abbreviated as MT-RGAN-H architecture) to construct prediction models on the solute parameters including excess molar refraction, dipolarity/polarizability, H-bond acidity (A), H-bond basicity (B), and logarithmic hexadecane–air partition coefficient. The resulting MT-RGAN-H model was proved to outperform single-task (ST) machine learning models, ST-RGAN models, MT-RGAN model, and previous models on the solute parameters. Importantly, the MT-RGAN-H model improved the prediction accuracy for A (5.6% increase) and B (4.3% increase) over the previous models. The MT-RGAN-H architecture also enables learning from multiple interrelated end points, even from end points with limited labeled cases, overcoming challenges of data insufficiency in some tasks. By visualizing the attention weights, the MT-RGAN-H model was well interpreted. The predicted solute parameters were further employed to predict six physicochemical parameters of chemicals, achieving better prediction accuracy over the previous optimal models. Therefore, this study provides an integrative “end-to-end” prediction scheme for the solute parameters, laying a foundation for accurately predicting the environmental partition behavior of chemicals. The MT-RGAN-H architecture can be further applied to construct prediction models on other interrelated end points, supporting sound management of chemicals.
A multitask relational graph convolutional network model with explicit hydrogen atoms and attention weights was developed for accurately predicting solute parameters of chemicals.
Improving Prediction Performance on Solute Parameters Using Multitask Relational Graph Convolutional Networks with Explicit Hydrogens
https://orcid.org/0009-0002-2696-5727
https://orcid.org/0009-0007-3724-5182
https://orcid.org/0000-0002-5756-3336
This study proposed multitask (MT) learning coupled with relational graph convolutional networks with attention weights (RGAN) and explicit hydrogens (abbreviated as MT-RGAN-H architecture) to construct prediction models on the solute parameters including excess molar refraction, dipolarity/polarizability, H-bond acidity (A), H-bond basicity (B), and logarithmic hexadecane–air partition coefficient. The resulting MT-RGAN-H model was proved to outperform single-task (ST) machine learning models, ST-RGAN models, MT-RGAN model, and previous models on the solute parameters. Importantly, the MT-RGAN-H model improved the prediction accuracy for A (5.6% increase) and B (4.3% increase) over the previous models. The MT-RGAN-H architecture also enables learning from multiple interrelated end points, even from end points with limited labeled cases, overcoming challenges of data insufficiency in some tasks. By visualizing the attention weights, the MT-RGAN-H model was well interpreted. The predicted solute parameters were further employed to predict six physicochemical parameters of chemicals, achieving better prediction accuracy over the previous optimal models. Therefore, this study provides an integrative “end-to-end” prediction scheme for the solute parameters, laying a foundation for accurately predicting the environmental partition behavior of chemicals. The MT-RGAN-H architecture can be further applied to construct prediction models on other interrelated end points, supporting sound management of chemicals.
A multitask relational graph convolutional network model with explicit hydrogen atoms and attention weights was developed for accurately predicting solute parameters of chemicals.
Improving Prediction Performance on Solute Parameters Using Multitask Relational Graph Convolutional Networks with Explicit Hydrogens
Xiao, Zijun (author) / Zhu, Minghua (author) / Chen, Jingwen (author)
ACS ES&T Water ; 4 ; 5969-5979
2024-12-13
Article (Journal)
Electronic Resource
English
Graph-Based Feature Matching Using Descriptive and Relational Parameters
| Online Contents | 1994
Rockburst Prediction via Multiscale Graph Convolutional Neural Network
| Springer Verlag | 2025