Prediction of pore properties of hierarchical porous silica templated on natural rubber

• Hierarchical porous silica (HPS) templated on natural rubber is made via sol-gel.
• HPS pore properties are satisfactorily predicted using ANN models (superior to CCD).
• A single hidden layer in ANN can predict HPS pore volume and BET surface area.
• Using a separate set of data as part of ANN training can identify model overfitting.

Hierarchical porous silica (HPS) templated on natural rubber (NR) was made via sol-gel technique using sodium silicate as a silica source. Macropores and mesopores with an ink-bottle type of morphology were randomly distributed in the HPS. Prediction of surface area and pore volume of the produced HPS was undertaken by central composite design (CCD) and artificial neural network (ANN) models, which were separately developed and which made use of the following variables: pH, calcination temperature, NR amount, and salt concentration. To obtain optimum feedforward back propagation networks, the number of hidden layers and the number of neurons in each hidden layer were varied. The use of an extra dataset in developing the optimum ANN architecture in each training cycle helped to easily locate when and where the overfitting phenomenon occurred, and thus the generalization performance of a network was not compromised. The quadratic polynomial models obtained using CCD poorly predicted surface area and pore volume (R2 < 0.65). In contrast, the optimum ANN models with a single hidden layer predicted both pore properties exceptionally well (R2 > 0.95).

Figure optionsDownload as PowerPoint slide