Research on the Prediction of Activated Carbon Properties Using Machine Learning

抄録

This study presents a predictive model utilizing artificial neural networks (ANNs), specifically Multilayer Perceptron (MLP) and Deep Neural Networks (DNN), to forecast the properties of activated carbon under various experimental conditions. The prediction model focuses on the Brunauer-Emmett-Teller (BET) surface area and total pore volume of the activated carbon, with the carbonization temperature, activation temperature, and activation agent as the determining factors. A dataset comprising around 100 samples was used for training the model and testing its accuracy. Results indicate that the DNN model, despite its increased complexity, exhibits superior performance over the MLP model in predicting the properties of activated carbon. The DNN model showed a Mean Absolute Percentage Error (MAPE) of 14.19% for the BET surface area prediction and 17.90% for the total pore volume prediction. The findings underscore the potential of using DNN models in optimizing activated carbon production processes and tailoring its properties for specific applications. Nonetheless, the study suggests the need for expanding the dataset and including more influential factors to further enhance the model's accuracy and reliability.