Predicting initial dissolution rates using structural features from molecular dynamics simulations

Abstract

Predicting the chemical durability of glass materials is important for various applications from daily life such as cell phone screens and kitchenware to advanced technologies such as nuclear waste disposal and biomedicine. In this work, we explored the prediction of the initial glass dissolution rates using structural features from molecular dynamics (MD) simulations for a series of glass compositions (total 28), including ZrO 2 -and V 2 O 5 -containing boroaluminosilicate, borosilicate, and aluminosilicate glasses. The initial dissolution rates (r 0 ) measured experimentally at 90 • C with varying solution conditions were correlated with structural features (e.g., polyhedral linkages and non-bridging oxygen species) obtained from MD simulations, either from this study or from literature. As hydrolysis of the glass network through breaking of the network former linkages (e.g., Si-O-Si, Si-O-Al, etc.) is a critical step of network glass dissolution, the statistics of these linkages obtained from MD were also correlated to r 0 through linear regression, where the coefficients of determination (R 2 ) and root mean square error are found to be 0.949 and 0.681, respectively. This model was compared and discussed with existing models developed by various approaches, including machine learning, the kinetic rate equation, topological constraint theory, and other descriptors from MD simulations. The discussion provides insights on future model improvements to predict glass dissolution. In addition, as the effect of V 2 O 5 on glass dissolution was not well studied comparing to ZrO 2 , the impact of V 2 O 5 was emphasized in this paper, suggesting that the impact is not the same across all glass compositions and test conditions.

Glass formulation and composition optimization with property models: A review

“…139 MD-based QSPR is mostly limited by the potential parameter developments, as well as the lack of bond energy information, as discussed previously. 60 Despite the ability of MD to simulate glasses containing oxides with multiple oxidation states (as demonstrated in studies on vanadium 120,130 and iron 195 ), the knowledge of the quantity of each oxidation state through experimental input is necessary. Furthermore, generating MD simulated structures of hundreds or thousands of glasses with a consistent simulation process requires significant computing resources, as it was found that system size and cooling rate have impacts on final simulated structures.…”

“…The initial dissolution rate (r 0 ) model, shown in Figure 6, used both pH and structural features obtained from MD, whereas using glass compositions alone was insufficient to represent measured data. 60 With these many options available as model features, feature selection analysis is the next important aspect in terms of model fitting. SHapley Additive exPlanations (SHAP) is a model-agnostic method for explaining individual predictors of ML models.…”

“…Linear regression can easily be fitted with multiple software such as Excel and OriginPro (OriginLab Corporation). Other common tools and software used in glass property model fitting, including, but not limited to, MATLAB (MathWorks), JMP (SAS Institute) (e.g., waste glass properties 29,60 and dissolution of phosphate glasses 63 ), Python scikit-learn 165 (e.g., predicting glass transition temperature 67,168 ), TensorFlow, 169 GPflow, 170 and PyTorch. 171…”

“…In addition, property models can include sample preparation details (e.g., heat treatment and redox) and/or testing conditions (e.g., temperature and pH) as model terms, as a few properties, such as chemical durability, 160 bioactivity, 161,162 and hardness, 163 can be highly dependent on thermal history and testing conditions. The initial dissolution rate ( r 0 ) model, shown in Figure 6, used both pH and structural features obtained from MD, whereas using glass compositions alone was insufficient to represent measured data 60 …”

“…(A) Predicted (error bars: 95% prediction interval) versus measured (error band: 30% r 0 ) plot of ln[ r 0 ] (g m −2 d −1 ) of a regression model fitted using pH and molecular dynamics (MD) structural information. (B) Predicted individual component effect of the structural model 60 …”