11:45 AM - 12:00 PM
[HDS11-10] Applicability of Activated Carbon for Emergency Response in Disasters and Accidents – Prediction of Adsorption Performance Using AI Machine Learning
Keywords:Prediction of Activated Carbon Adsorption Performance, Chemical Release, AI Machine Learning, Adsorption Coefficient (Log K), Risk Management in Disasters and Accidents, Emergency Response
Chemical release accidents in drinking water sources can lead to the suspension of water intake and water purification supply, necessitating the establishment of rapid removal techniques. Activated carbon, characterized by its porous structure and large surface area, is widely used as an adsorbent to remove chemical substances released into rivers during emergency incidents. However, its adsorption capacity depends on molecular structure and pore characteristics, and comprehensive data for all chemical substances remain insufficient. Conducting individual adsorption experiments for chemical substances lacking experimental adsorption data is both time-consuming and costly, which may hinder immediate and appropriate responses to future release incidents caused by natural disasters. Furthermore, the molecular structure of activated carbon is irregular, making it challenging to achieve high-precision quantitative predictions using conventional physicochemical models.
Therefore, this study aims to develop and evaluate an AI model capable of predicting the adsorption coefficient (log K) of chemical substances onto activated carbon.
First, a deep learning model utilizing a graph convolutional network (GCN) was constructed using the machine learning software M-EVO (HPC Systems Inc.). The dataset was based on experimental adsorption data reported by Abe (1986), selecting 95 compounds with well-defined molecular structures from 101 compounds available in the literature on activated carbon adsorption. These compounds were divided into a training set (85 compounds) and a test set (10 compounds), and the learning accuracy was evaluated in stages. To ensure the uniformity of physicochemical properties among the compounds used as training data, a random selection process was repeated five times within the 95 compounds. The model’s predictive accuracy was assessed by confirming that the adsorption coefficients of the 10 test compounds were well predicted compared to literature values (Abe, 1986), achieving an R² value of 0.895.
Using this trained model, 11 substances were selected from the experimental dataset of the Environmental Protection Agency (EPA, 1980) under experimental conditions similar to those of Abe (1986), and their adsorption coefficients onto activated carbon were predicted. Additionally, the predicted adsorption coefficients were compared with literature values from EPA (1980), and their correlation was analyzed by calculating the coefficient of determination (R²) and examining the error distribution.
As a result, the R² value between the literature values and the model’s predicted adsorption coefficients (log K) was 0.902, indicating a strong correlation. The root mean square deviation (RMSD) was 0.534, demonstrating that the predicted values were generally consistent with the literature values. Among the aromatic hydrocarbons, benzene, toluene, p-xylene, styrene, and ethylbenzene were well reproduced. However, among the aromatic compounds, 4,4'-methylene-bis (literature value: 2.28, predicted value: 3.43), 4-aminobiphenyl (literature value: 2.30, predicted value: 3.14), and naphthalene (literature value: 2.12, predicted value: 2.81) showed a tendency for the predicted adsorption coefficients to be overestimated by approximately 0.7 to 1.2 compared to the experimental values.
Possible reasons for this discrepancy include the presence of experimental data in the training set that deviate from other reported values and the fact that while the GCN model effectively learned local molecular structures, it may have only partially accounted for steric size and interaction effects. Future work should involve conducting new experiments to verify the consistency between predicted and measured adsorption coefficients. Additionally, improvements in feature selection and dataset expansion will be pursued to develop a more accurate predictive model.
Therefore, this study aims to develop and evaluate an AI model capable of predicting the adsorption coefficient (log K) of chemical substances onto activated carbon.
First, a deep learning model utilizing a graph convolutional network (GCN) was constructed using the machine learning software M-EVO (HPC Systems Inc.). The dataset was based on experimental adsorption data reported by Abe (1986), selecting 95 compounds with well-defined molecular structures from 101 compounds available in the literature on activated carbon adsorption. These compounds were divided into a training set (85 compounds) and a test set (10 compounds), and the learning accuracy was evaluated in stages. To ensure the uniformity of physicochemical properties among the compounds used as training data, a random selection process was repeated five times within the 95 compounds. The model’s predictive accuracy was assessed by confirming that the adsorption coefficients of the 10 test compounds were well predicted compared to literature values (Abe, 1986), achieving an R² value of 0.895.
Using this trained model, 11 substances were selected from the experimental dataset of the Environmental Protection Agency (EPA, 1980) under experimental conditions similar to those of Abe (1986), and their adsorption coefficients onto activated carbon were predicted. Additionally, the predicted adsorption coefficients were compared with literature values from EPA (1980), and their correlation was analyzed by calculating the coefficient of determination (R²) and examining the error distribution.
As a result, the R² value between the literature values and the model’s predicted adsorption coefficients (log K) was 0.902, indicating a strong correlation. The root mean square deviation (RMSD) was 0.534, demonstrating that the predicted values were generally consistent with the literature values. Among the aromatic hydrocarbons, benzene, toluene, p-xylene, styrene, and ethylbenzene were well reproduced. However, among the aromatic compounds, 4,4'-methylene-bis (literature value: 2.28, predicted value: 3.43), 4-aminobiphenyl (literature value: 2.30, predicted value: 3.14), and naphthalene (literature value: 2.12, predicted value: 2.81) showed a tendency for the predicted adsorption coefficients to be overestimated by approximately 0.7 to 1.2 compared to the experimental values.
Possible reasons for this discrepancy include the presence of experimental data in the training set that deviate from other reported values and the fact that while the GCN model effectively learned local molecular structures, it may have only partially accounted for steric size and interaction effects. Future work should involve conducting new experiments to verify the consistency between predicted and measured adsorption coefficients. Additionally, improvements in feature selection and dataset expansion will be pursued to develop a more accurate predictive model.