Machine learning modeling of the atomic structure and physical properties of alkali and alkaline-earth aluminosilicate glasses and melts
arXiv:2304.12123 · doi:10.1016/j.jnoncrysol.2023.122481
Abstract
The first version of the machine learning greybox model i-Melt was trained to predict latent and observed properties of KO-NaO-AlO-SiO melts and glasses. Here, we extend the model compositional range, which now allows accurate predictions of properties for glass-forming melts in the CaO-MgO-KO-NaO-AlO-SiO system, including melt viscosity (accuracy equal or better than 0.4 log Pas in the 10-10 log Pas range), configurational entropy at glass transition ( 1 J mol K), liquidus ( 60 K) and glass transition ( 16 K) temperatures, heat capacity ( 3 \%) as well as glass density ( 0.02 g cm), optical refractive index ( 0.006), Abbe number ( 4), elastic modulus ( 6 GPa), coefficient of thermal expansion ( 1.1 10 K) and Raman spectra ( 25 \%). Uncertainties on predictions also are now provided. The model offers new possibilities to explore how melt/glass properties change with composition and atomic structure.
arXiv commentAccepted version. Changes were made following reviews. New figures (4, 5, 6) were added. Text in section 3.1 has been modified accordingly. Figures 7 and 8 were improved. Small modifications of the text were made to correct typos, to add references or to clarify the text. Two lines were added in Table 1 to document errors on glass transition temperature and entropy
References in corpus (6)
- Scikit-learn: Machine Learning in Python
- Gaussian Error Linear Units (GELUs)
- Tune: A Research Platform for Distributed Model Selection and Training
- Signature of small rings in the Raman spectra of normal and compressed amorphous silica: A combined classical and ab initio study
- ViscNet: Neural network for predicting the fragility index and the temperature-dependency of viscosity
- Reliable Prediction Errors for Deep Neural Networks Using Test-Time Dropout