09:45 〜 10:00
[SIT15-04] Predicting the viscosity of aluminosilicate melts with machine learning
キーワード：magma, volcanism, viscosity, machine learning
Aluminosilicate melts are key to understand volcanism and magmatic processes. Their rapid quench further allows the formation of glasses used in different industrial and technological applications. Among important properties, viscosity is central as it controls the way those melts flow. It varies by decades with temperature, pressure and melt composition, including volatile elements. Models exist but are subject to different limitations. For instance, they only target a specific compositional set, such as geological melts. They also usually offer no valid idea of their uncertainties.
In recent years, the rise of machine learning led to the proposition of several models leveraging this method to predict the viscosity (and other properties) of geologic or industrial melts. Among tested algorithms, artificial neural networks (ANN) found uses in different models. Different implementations exist. For instance, given a melt composition, blackbox ANN models predict temperature at different fixed viscosity values; the general temperature dependence can then be retrieved via a second least-squared fit of the generated values, using an equation such as MYEGA or VFT. Greybox ANN models propose to embed those equations within their algorithm. It is not clear which methods yield best results, and to which extend different machine learning algorithms are robust for implementing them. Moreover, evaluation of the uncertainties affecting model predictions is not straightforward.
In this study, we present a new database of experimental viscosity measurements of melts in the SiO2-TiO2-Al2O3-Fe2O3-FeO-MnO-Na2O-K2O-MgO-CaO-P2O5-H2O system. This database is composed of 15,919 data points from 2855 different compositions, ranging from pure silica to geological and industrial melts. Using this database, we show how different machine learning algorithms perform. We discuss the possibility of retrieving robust predictive uncertainties, and demonstrate that models embedding domain knowledge outperforms existing models, with predictive root-mean-squared-errors equal to, or lower than 0.4 log Pa·s.
In recent years, the rise of machine learning led to the proposition of several models leveraging this method to predict the viscosity (and other properties) of geologic or industrial melts. Among tested algorithms, artificial neural networks (ANN) found uses in different models. Different implementations exist. For instance, given a melt composition, blackbox ANN models predict temperature at different fixed viscosity values; the general temperature dependence can then be retrieved via a second least-squared fit of the generated values, using an equation such as MYEGA or VFT. Greybox ANN models propose to embed those equations within their algorithm. It is not clear which methods yield best results, and to which extend different machine learning algorithms are robust for implementing them. Moreover, evaluation of the uncertainties affecting model predictions is not straightforward.
In this study, we present a new database of experimental viscosity measurements of melts in the SiO2-TiO2-Al2O3-Fe2O3-FeO-MnO-Na2O-K2O-MgO-CaO-P2O5-H2O system. This database is composed of 15,919 data points from 2855 different compositions, ranging from pure silica to geological and industrial melts. Using this database, we show how different machine learning algorithms perform. We discuss the possibility of retrieving robust predictive uncertainties, and demonstrate that models embedding domain knowledge outperforms existing models, with predictive root-mean-squared-errors equal to, or lower than 0.4 log Pa·s.