<p>In this work, we developed a multi-output machine learning pipeline to jointly predict glass-forming region (GFR) tendency and radiation shielding properties across varying energy levels in binary oxide glass systems. A dataset of 2284 compositions spanning 16 binary systems was prepared, and shielding properties were considered at 25 different photon energies (0.015-15 MeV). Compositions were curated from SciGlass-reported binary oxide glasses, and the energy-dependent attenuation descriptors were computed using XCOM-based photon interaction data over 0.015-15 MeV. Using 9 different composition features, we compared a linear Ridge regression baseline with a multi-layer-perceptron (MLP) model to predict 151 targets, including a GFR confidence score (0-100), linear attenuation coefficient (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\:\mu\:\)</EquationSource> </InlineEquation>), mass attenuation coefficient (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\:\mu\:/\rho\:\)</EquationSource> </InlineEquation>), mass energy-absorption coefficient (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\:{\mu\:}_{en}/\rho\:\)</EquationSource> </InlineEquation>), half-value layer (HVL), tenth-value layer (TVL), and mean free path (MFP). On a held-out test set of 457 compositions, the MLP model achieved an overall performance of R² = 0.9906 across the shielding targets, outperforming the Ridge model performance of R² = 0.9732. The MLP model showed strong performance for GFR prediction with R² = 0.871, with a wide margin of accuracy compared to the Ridge model performance of R² = 0.267. On the other hand, attenuation-related targets remained highly accurate across models, with <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\:\mu\:/\rho\:\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\:{\mu\:}_{en}/\rho\:\)</EquationSource> </InlineEquation> predicted at <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\:{R}^{2}\approx\:1.000\)</EquationSource> </InlineEquation>, while <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\:\mu\:\)</EquationSource> </InlineEquation>, HVL, TVL, and MFP were predicted at <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(\:{R}^{2}\approx\:0.989\)</EquationSource> </InlineEquation>. Finally, we demonstrated how the trained surrogate can generate GFR maps and enable rapid composition screening at 1 MeV. Among the screened systems at 1 MeV, the lowest predicted HVL values were obtained for SiO₂-PbO (3/97 wt%, HVL = 1.277 cm, GFR confidence score = 67.93) and B₂O₃-Bi₂O₃ (6/94 wt%, HVL = 1.350 cm, GFR confidence score = 85.80). Overall, this study provides a practical modeling route that supports efficient identification of binary oxide glasses balancing formability and shielding performance and guides future studies toward uncertainty quantification and experimental benchmarking in broader compositional spaces.</p>

Machine learning prediction of glass-forming region and radiation shielding properties in binary oxide glass systems

期刊论文
发表日期：

2026年6月22日

查看全文

M. S. Al-Buriahi,
M. U. Baskin,
Jamila S. Alzahrani,
Khadijah Mohammedsaleh Katubi

摘要

In this work, we developed a multi-output machine learning pipeline to jointly predict glass-forming region (GFR) tendency and radiation shielding properties across varying energy levels in binary oxide glass systems. A dataset of 2284 compositions spanning 16 binary systems was prepared, and shielding properties were considered at 25 different photon energies (0.015-15 MeV). Compositions were curated from SciGlass-reported binary oxide glasses, and the energy-dependent attenuation descriptors were computed using XCOM-based photon interaction data over 0.015-15 MeV. Using 9 different composition features, we compared a linear Ridge regression baseline with a multi-layer-perceptron (MLP) model to predict 151 targets, including a GFR confidence score (0-100), linear attenuation coefficient ( \(\:\mu\:\) ), mass attenuation coefficient ( \(\:\mu\:/\rho\:\) ), mass energy-absorption coefficient ( \(\:{\mu\:}_{en}/\rho\:\) ), half-value layer (HVL), tenth-value layer (TVL), and mean free path (MFP). On a held-out test set of 457 compositions, the MLP model achieved an overall performance of R² = 0.9906 across the shielding targets, outperforming the Ridge model performance of R² = 0.9732. The MLP model showed strong performance for GFR prediction with R² = 0.871, with a wide margin of accuracy compared to the Ridge model performance of R² = 0.267. On the other hand, attenuation-related targets remained highly accurate across models, with \(\:\mu\:/\rho\:\) and \(\:{\mu\:}_{en}/\rho\:\) predicted at \(\:{R}^{2}\approx\:1.000\) , while \(\:\mu\:\) , HVL, TVL, and MFP were predicted at \(\:{R}^{2}\approx\:0.989\) . Finally, we demonstrated how the trained surrogate can generate GFR maps and enable rapid composition screening at 1 MeV. Among the screened systems at 1 MeV, the lowest predicted HVL values were obtained for SiO₂-PbO (3/97 wt%, HVL = 1.277 cm, GFR confidence score = 67.93) and B₂O₃-Bi₂O₃ (6/94 wt%, HVL = 1.350 cm, GFR confidence score = 85.80). Overall, this study provides a practical modeling route that supports efficient identification of binary oxide glasses balancing formability and shielding performance and guides future studies toward uncertainty quantification and experimental benchmarking in broader compositional spaces.