Chlorine-containing silanes such as SiH
3Cl and SiH
2Cl
2 have been widely used to produce the high-purity polycrystalline silicon that plays an essential role in semiconductors and photovoltaic solar cells. The demand for semiconductors is particularly high and so large quantities of high-purity polycrystalline silicon are required. During the silicon manufacturing process, the formation of high-purity polycrystalline silicon from purified chlorine-containing silanes generates byproducts consisting of silicon, silane compounds, chlorine, hydrogen and oxygen as off-gases. The silane compounds in turn comprise silanes, silanols containing a variety of elements. Importantly, silane compounds in the off-gases are highly reactive and can cause explosions. As an example, an explosion occurred at a high-purity polycrystalline silicon production facility in Japan, in 2014, resulting in five fatalities and 13 injuries. Several studies concerning the hazards of silane compounds have been conducted to date, and various explosive silanes have been proposed. However, hazardous reactions and explosive silane compounds have not been determined yet, leading to a lack of information to help design a safer process. An improved understanding of the hazards of these proposed silane compounds will require an investigation of the thermochemical properties of silane compounds along with data such as heat of combustion and decomposition. Unfortunately, it is very difficult to obtain experimental heat of formation for silane compounds due to technical difficulties associated with the high reactivity of silane compounds. To obtain heat of formation for silane compounds, computational methods are effective, and there are two computational methods to obtain the heat of formation: quantum chemical calculations and group additivity methods. Quantum chemical calculations provide accurate estimates of the heat of formation, but their computational cost for large silane compounds is too high. In the group additivity methods, the heat of formation consists of the contribution of a group of molecules, i.e., the heat of formation is estimated by adding up the group additivity values (GAVs), and its computational cost is negligible, which is why these methods are suitable for obtaining thermodynamic data for silane compounds. In previous works, group additivity methods have been applied to silane compounds containing Si, H, and O. However, group additivity methods have not been applied to silane compounds containing a variety of elements, which means that it’s impossible to obtain thermodynamic data of various silane compounds in the off-gases from the silicon manufacturing process. Therefore, the purpose of this study is to try to apply the group additivity methods to silane compounds containing a variety of elements. In addition, the use of machine learning in the obtainment of GAVs facilitates the application of the group additivity methods, and the accumulation of the heat of formation of silane compounds leads to improved accuracy. As the first step, we try to apply the group additivity methods to silane compounds containing Si, H, O, and Cl.
This study consists of four steps. The first step was to construct a database including thermodynamic data and groups of small (Si<5) silane compounds. Thermodynamic data for a total of 312 small silane compounds including linear and cyclic were obtained from quantum chemical calculations at the CBS-QB3//M06-L/6-311+G(3d,p) level of theory, and groups of these molecules were counted. The second step was to obtain GAVs using machine learning. The linear regression with the database constructed in the first step as training data was conducted using python 3.7.9, which provided GAVs of a total of 30 groups from the mathematical model. The third step was to validate the GAVs, where the heat of formation from quantum chemical calculations (ΔfH from QCC) used as training data and heat of formation estimated from GAVs (ΔfH from GAVs) were compared to determine the prediction accuracy of ΔfH from GAVs. As a result, ΔfH from GAVs showed good agreement with ΔfH from QCC, and R-squared (R2) score between ΔfH from QCC and ΔfH from GAVs was 0.9999. The fourth step was to estimate thermodynamic data for silane compounds using GAVs obtained in this study. ΔfH of silane polymers that had not previously existed were obtained using the group additivity methods, which leads to further understanding of the reactivity and hazards of silane compounds in the off-gases from the silicon manufacturing process.
