Glass dissolution resulting from the dissolution of soluble elements from the glass surface forms the alteration layer. It is known that the chemical structure and volume of the alteration layer vary with different dissolution conditions, such as solution and glass compositions and time. In this study, we study a model to estimate the alteration layers volume via experimental and modeling approaches aided by a machine-learning.
The glass dissolution tests were performed for the 20 glasses with various compositions.The pristine glasses has the composition based on ISG1 shown in Image 1 and are prepared by the customary melt quenching method.The dissolution was conducted under static conditions at 140°C for seven days according to the standard for glass melting tests. The normalized mass loss (NL) of each element in the pristine glass was calculated from the elemental analysis of the leaching solution. The coordination structures of Si and B for pristine and altered glasses were determined from solid state 11B and 29Si NMR spectroscopy.
The alteration layer volume was calculated from the leaching amount of B. The data showed the nonlinear behavior for Na content in glass compositions, meaning that the simple ion exchange model between H3O+ and Na could not explain the formation volume of the alteration layer.
The correlation matrix for all compositions and leaching elements was computed to investigate the correlations. The highest correlation in NLs was found for Na-Si, and Na-B. Solid-state 29Si NMR spectra showed several peaks with broadened line shapes, which were assigned to the Qn structure involving neighbor AlO4. Due to the broad line shape, it is not easy to quantify these chemical species by deconvolution. The deconvolution was roughly performed using two Gaussian functions assuming Q3 and Q4 or Q4(1Al, 2Al) structures.
Solid-state 11B NMR spectra were also deconvoluted by considering second-order quadrupolar interaction. Four peaks corresponding to ring and non-ring 3-coordinate B and two 4-coordinate B species were assumed. A series of the experimental dataset for atomic species derived from solid-state NMR were used to calculate correlation the correlation matrix. The correlation matrix showed that the Q4 population has a negative correlation with Al NL (correlation coefficient: -0.66). Finally, a prediction model of B NL from compositions, the population of atomic species, and pristine glass density were built using linear regression.