News

Historically, new glasses were developed through slow, costly experimentation guided by experts’ intuition. This is changing because ML algorithms can now use large datasets to uncover composition–property relationships across vast compositional spaces. A key stimulus for this shift was the 2004 question posed by Zanotto and Coutinho: how many non-crystalline solids can be made from all elements of the periodic table? They showed that over 1052 glass compositions are theoretically possible, yet only ~106 have ever been synthesized and characterized [2] — an enormous, largely unexplored "glass genome" that human effort alone could never cover. Because glass properties depend mainly on composition, ML is a promising tool to conquer this space at a scale beyond human intuition.

Early ML studies were limited by small datasets (<1,000 points). A major breakthrough came when Cassar et al. (2018) trained neural networks on ~55,000 compositions to predict glass transition temperature (Tg), outperforming classical empirical equations and demonstrating the value of large databases such as SciGlass and INTERGLAD [3]. This opened the door to predicting other properties (liquidus temperature, elastic modulus, refractive index, etc) [4]. Cassar et al. (2021a) broadened the scope by testing multiple algorithms and showing that ML can not only predict oxide-glass properties but also provide valuable insights into them. Using SHAP, they revealed which oxides most strongly influence each property [4], countering the "black box" criticism. Building on this interpretability, the authors' recent work applies SHAP to quantify a chemical and property "distance" between glass compositions [5], helping identify unique glasses and assess how novel an ML-designed glass is. Extending beyond oxides, Mastelini et al. (2022) showed ML can also uncover subtle composition–property relationships in chalcogenide glasses [6]. Another trend is embedding physical knowledge into ML models. Mannan et al. (2023) combined symbolic reasoning with ML to model glass hardness, incorporating physically interpretable descriptors and load effects rather than relying solely on neural-network regression [7]. This points toward hybrid strategies merging data-driven learning with thermodynamics, structural chemistry, and domain expertise.

Beyond prediction, ML workflows must now enable inverse design: proposing new glasses tailored to achieve desired properties. Cassar et al. (2021b) coupled ML with genetic algorithms to design optical glasses optimized simultaneously for refractive index and Tg [8]. Souza et al. (2026) used ML-assisted design to develop lead-free niobium "crystal" glasses with improved optical performance, addressing environmental concerns [9]. Such approaches can screen millions of candidate compositions to identify promising ones before experimental synthesis.

The next challenge is data itself. Most of the literature still relies on the same two databases, SciGlass and INTERGLAD, which cannot sustain the field's growth alone. Ravinder et al. (2021) already flagged automated extraction of glass data from the literature as one of 21 key challenges facing artificial intelligence in glass science [10]. The authors' own study, Thomaello et al. (2025), addresses this gap by converting patent tables into structured composition–property datasets [11], illustrating how automated data-extraction pipelines can unlock knowledge still buried in the literature.

For young researchers, the lessons are clear: data handling and statistical modelling have become core scientific skills; ML complements rather than replaces fundamental glass science; and opportunities span photonics, healthcare, energy storage, nuclear waste immobilization, bioactive implants, high-strength displays, and quantum technologies. Persistent challenges include scarce datasets for many properties, sampling biases, limited treatment of thermal/processing history, and unreliable extrapolation beyond known compositions, requiring collaboration among experimentalists, computational scientists, statisticians, and AI specialists. The convergence of glass science and ML remains in its early stages but is already reshaping how researchers understand, predict, and design disordered materials.

Acknowledgments

The ML work has been doable due to invaluable collaboration with Daniel Cassar, André Ponce de Leon and several other co-authors. In the past year, we have had weekly meetings with other ML geeks: Thiago Ramos, Daiane Zuanetti, and Thomaz Aguena. 

Disclaimer: Several other groups in the world are working on these problems, we used our own work as examples of current developments.

References

[1] Zanotto, E.D. Artificial intelligence is reshaping glass understanding and design. ACerS Bulletin, vol.105, n.6, pg.3, Aug. 2026 

[2] Zanotto, E.D. and Coutinho, F.A.B. "How many non-crystalline solids can be made from all the elements of the periodic table?", J. Non-Cryst. Solids 2004, v.347, pp.285-288.

[3] Cassar, D.R., de Carvalho, A.C.P.L.F., Zanotto, E.D. Predicting glass transition temperatures using neural networks, Acta Materialia, 2018, v. 159, pp. 249-256. 

[4] Cassar, D.R., Mastelini, S.M., Botari, T., Alcobaça, E., de Carvalho, A.C.P.L.F., Zanotto, E.D. Predicting and interpreting oxide glass properties by machine learning using large datasets. Ceramics International, 2021a, v. 47, n. 17, pp. 23958–23972.

[5] Rosante, R.B., Thomaello, G.L., Aguena, T.Y.B., Ramos, T.R., Zuanetti, D.A., Zanotto, E.D. Quantifying oxide glass similarity through learned distances: a SHAP-based Earth Mover’s Distance. Journal of Non-Crystalline Solids, 2026, v. 691, 124303

[6] Mastelini, S.M.; Cassar, D.R.; Alcobaca, E.; Botari, T.; de Carvalho, A.C.P.L.F; Zanotto, E.D. Machine learning unveils composition-property relationships in chalcogenide glasses. Acta Materialia, 2022, v. 240, 118302. 

[7] Mannan, S., Zaki, M., Bishnoi, S., Cassar, D.R., Jiusti, J., Faria, J.C.F., Christensen, J.F.S., Gosvami, N.N., Smedskjaer, M.M., Zanotto, E.D., Krishnan, N.M.A. Glass Hardness: Predicting Composition and Load Effects via Symbolic Reasoning-Informed Machine Learning. Acta Materialia, 2023, v.255, 119046

[8] Cassar, D.R., dos Santos, G.G., Zanotto, E.D. Designing optical glasses by machine learning coupled with a genetic algorithm, Ceramics International, 2021, v.47, n. 8, pp.10555 - 10564. 

[9] Souza, M.L., dos Santos, G.G., Cassar, D.R., Zanotto, E.D. Machine-learning assisted design of lead-free niobium crystal glass with improved optical properties. Ceramics International. 2026, v.52, n. 11, Part A, pp.16129-16142

[10] Ravinder; Venugopal, V.; Bishnoi, S.; Singh, S.; Zaki, M.; Grover, H. S.; Bauchy, M.; Agarwal, M.; Krishnan, N. M. A. Artificial intelligence and machine learning in glass science and technology: 21 challenges for the 21st century. Int. J. Appl. Glass Sci., 2021, v. 12, n. 3, pp. 277–292. 

[11] Thomaello, G. L.; Aguena, T. Y. B.; Costa, E. T.; Rosante, R. B.; Ramos, T. R.; Zuanetti, D. A.; Zanotto, E. D.  From Patents to Dataset: Scraping for Oxide Glass Compositions and Properties. arXiv, 2025, arXiv:2511.16366.

----------------------------------------------------------------------------------------------------------

Personal note by Prof. E. D. Zanotto

This line of research was sparked in 2017 by a question from an engineer at an American company who was studying my 2004 paper (ref. [2]). He wondered how anyone could explore a 1052 compositional space in a reasonable timeframe, proposing high-throughput experimentation. Instead, I had an insight: why not use AI? Although I knew virtually nothing about artificial intelligence at the time, I pitched the idea to my PhD student, Daniel Cassar, who was passionate about coding and simulations. We brought in AI expert, Prof. André Ponce de Leon, and our research journey officially began!

Edgar Dutra Zanotto 

edz@ufscar.br

LinkedIn Profile

Gustavo Laranja Thomaello (student)

gustavothomaello@estudante.ufscar.br

LinkedIn Profile

Rafael Baságlia Rosante (student)

rbrosante@estudante.ufscar.br

LinkedIn Profile

Department of Materials Engineering, Federal University of São Carlos (UFSCar), Brazil