From Summation to AI-Assisted Data-Driven Methods of Glass Property Calculations

Glass is widely used in various fields like smartphone screens, optical fibers and architectural windows. Nevertheless, the methods for calculating glass properties are explored. The amorphous structure of glass presents unique challenges for predicting its properties, compared to crystalline materials. This review introduces the evolution of methods used to calculate glass properties, from conventional summation techniques to contemporary AI-assisted data-driven approaches, and represents recent progress and future direction of this field.

A foundational work was performed by Winkelman and Schott in 1894, who introduced early summation methods for predicting glass properties, such as the density, thermal expansion coefficient, and refractive index. Despite the non-crystalline nature of glass, some pioneers laid the groundwork for further advancements, i.e., composition-property additivity methods developed. These methods are rudimentary but crucial for understanding a relationship between glass composition and its properties. As research progressed, structure-based methods emerged as a significant leap forward. The methods, including the structural unit model, phase diagram approach, and topological constraint theory, incorporate the microstructural information. For instance, the phase diagrams visualize how different components interact within glass formulations, thus predicting critical glass properties such as the melting point and the viscosity. This is instrumental in controlling production processes and improving manufacturing outcomes. Besides, the development of topological constraint theory (TCT) is another milestone. The TCT method provides insights into atomic-scale constraints within amorphous materials, explaining why certain glass products exhibit specific behaviors., The material responses are predicted under various conditions without getting bogged down in complex jargon via understanding the arrangement of atoms. While these approaches enhance the understanding of composition-structure-property relationships, they are primarily applicable to simple systems due to limited structural data for complex multi-component glasses.

The integration of information technology and materials science leads to another significant advancements in glass research, complementing conventional experimental methods with the development of glass composition-property databases and simulation software. These tools enable efficient and cost-effective design and optimization of new functional glasses. Representative databases include SciGlass, INTERGLAD, GlassPy, and SMARTDATA, each offering unique features in data scale, prediction capabilities, machine learning integration, and industrial applications. Structural information of glass materials and their properties (i.e., thermal, optical, chemical, and mechanical properties) are collected in these databases with integrated statistical analysis and machine learning functionalities. In addition, specialized databases also cater to specific applications. For instance, the databases of nuclear waste vitrification are constructed in different countries like USA, South Korea, and China. Models are constructed to predict glass properties based on nuclear waste compositions, and machine learning-based design algorithms are adopted to optimize formulations for stability and safety of the nuclear waste glass in long-term storage.

Since glass properties are ultimately determined by structural changes induced by its composition, it is crucial to investigate the relationships among composition, structure, and properties. Two main modeling approaches emerge for designing glass, i.e., the statistical modeling with measured structural data, known as the Glass Structure Gene Simulation Method (GSgM), and computational methods such as ab initio molecular dynamics (AIMD), classical molecular dynamics (MD), and quantitative structure-property relationship (QSPR) models. The GSgM introduces structural data into conventional composition-property statistical models, enabling more accurate predictions via concerning composition, structure, and property relationships. Glass is considered as a network of structural units that statistically affect overall properties, and structural data can be integrated to obtain the Cornell s first-order statistical model. AIMD is continuously developed to overcome classical MD limitations. However, classical MD still remains a necessity for larger systems and longer timescales, despite relying on empirical potentials that approximate true potential energy surfaces. Moreover, recent progress on machine learning potentials and QSPR method offers promising alternatives, combining high accuracy with efficiency. MLPs use neural networks to fit potential functions from first-principles data, while QSPR links structural descriptors to material properties, enabling precise predictions. These integrated approaches aim to overcome computational limitations and improve glass property modeling, offering effective tools for materials design and optimization.

Furthermore, the integration of artificial intelligence (AI) and machine learning algorithms revolutionizes glass research via enabling the analysis of vast datasets to uncover patterns and correlations. This advancement improves the accuracy of predicting glass properties, facilitating the discovery of new glass formulations with desirable characteristics. Note that AI excels in identifying non-linear relationships that are challenging for conventional methods, opening up novel avenues for innovation in glass development. Despite these advancements, some challenges persist. The effectiveness of AI models is profoundly affected by the quality and comprehensiveness of the training data, which can be affected by factors such as instrumentation, experimental conditions, and human involvement. Ensuring high-quality datasets is therefore paramount to achieving reliable predictions in this complex field. Also, the integration of diverse computational techniques into a unified framework presents significant technical and logistical challenges. While AI offers powerful tools for advancing glass research, overcoming these integration hurdles remains essential for fully harnessing its potential.

The advancements in glass property calculations are shifting from data accumulation to high-efficiency prediction, with databases differing in scale, openness, algorithm integration, and applicability. Future developments will focus on data sharing, multi-method integration, and experimental-computational-machine learning synergies, providing an enhanced technical support for both scientific research and industrial applications of glass materials. Some methodologies can merge advances in materials science with cutting-edge computational tools. High-throughput preparation techniques promise a rapid discovery of new glass compositions via enabling swift synthesis and testing. Pairing these with advanced characterization methods like spectroscopy or diffraction can generate datasets for AI models, driving a further innovation.