Metallic glasses, created by quenching of molten alloys, lack crystalline structure resulting in distinctive property combinations for diverse potential applications, from golf clubs to transformer cores. However, glass formation in metals is poorly understood, greatly restricting the development of novel alloys. Many empirical rules and criteria aiming to describe the glass-forming ability of alloy compositions have emerged since metallic glasses were first discovered in 1959, but they have limited predictive power and do little to advance fundamental understanding. University of Cambridge’s Department of Materials Science & Metallurgy researchers RM Forrest & AL Greer have applied the fourth paradigm of scientific research by training a machine-learning model of glass-forming ability on a large dataset of experimental measurements, not merely to produce a predictive tool, but also to extract theoretical insights. While recovering several established empirical criteria, they identified an important relation between glass-forming ability and the electronic structures of the elements in an alloy. Novel alloy compositions with high glass-forming ability and also crystalline high-entropy alloys may be systematically designed by selecting elements to induce mutual changes in atomic radii.Metallic glasses are created from molten metal alloys by cooling at a rate fast enough to prevent any significant degree of crystallization. The lack of time for transition into a crystalline ordering leaves the atoms in a liquid-like structure.2 Alloy compositions for which it is possible to produce samples with a minimum thickness of 1–10 millimeters are referred to as bulk metallic glasses, while others may be obtainable only as thin glassy ribbons. The amorphous structures of metallic glasses give them interesting properties, with potential applications in many areas, from sporting equipment to aircraft and automotive components. The relative novelty of metallic glasses s means further glass-forming alloy compositions are awaiting discovery, with the promise of meeting previously inaccessible requirements for applications.The ability of an alloy composition to resist arrangement into an ordered crystalline phase upon cooling is referred to as the glass-forming ability. The glass-forming ability of an alloy composition may be directly quantified via the maximum achievable diameter of a rod cast into a fully glassy state, Dmax. Predicting the glass-forming ability of alloy compositions is a central goal in MG research. This research is plagued by empirical rules and trial-and-error experimentation, with frustratingly many proposed criteria for glass-forming ability, each often published with claims of superiority yet limited proven applicability. The present work explores the power of the burgeoning ‘fourth paradigm’ of scientific discovery, that being the utilization of data to train machine-learning models of physical phenomena, which in turn inform our theoretical understanding.Machine-learning involves the creation of models that can improve their performance via exposure to data. Materials science is no exception to the rapid spread of Machine-learning as a research tool. Machine-learning approaches to materials science use the vast amount of experimental data now available to model the physical laws governing observed phenomena. Neural networks, in particular, are commonly applied in Machine-learning and have been widely used within materials science to model atomic interactions, predict synthesis routes,15 reconstruct structures from imaging, identify phases and transitions, predict welding criteria, predict material properties and to predict the existence of novel materials, among others.Researchers said “Leveraging this insight and the predictive ability of the ensemble neural-network model, we make a number of suggestions for novel binary and ternary glass-forming alloy systems that will exhibit increases in the deviation of atomic radii due to the balancing of electron density. In the search for these alloy systems, lithium frequently appears due its combination of small radius and low nWS, and has been seen elsewhere in the literature as an addition to alloy compositions which increases glass-forming ability. We emphasize the need for more data to be published on glass-forming alloy compositions, both detailing the successful creation of glassy alloys but, importantly, also failed attempts. This is essential to avoid the study of BMGs becoming entrenched in well-trodden areas of composition-space, dependent on empirical rules. Further, the relative lack of BMG-forming alloy compositions with large Dmax presents a significant challenge to ML, as little information is provided from which to learn the rules that determine their existence. Nevertheless, the transferability of our model is successfully tested via k-folds cross-validation. Future work in this direction may consider more advanced theories than the Thomas–Fermi model for electron density, such as the Thomas–Fermi–Dirac model which includes the exchange energy, or a full density-functional theory treatment. In addition to nWS, we identify several other features to be important to the neural-network model, of which deeper investigation may return further useful insights into glass formation.”
Metallic glasses, created by quenching of molten alloys, lack crystalline structure resulting in distinctive property combinations for diverse potential applications, from golf clubs to transformer cores. However, glass formation in metals is poorly understood, greatly restricting the development of novel alloys. Many empirical rules and criteria aiming to describe the glass-forming ability of alloy compositions have emerged since metallic glasses were first discovered in 1959, but they have limited predictive power and do little to advance fundamental understanding. University of Cambridge’s Department of Materials Science & Metallurgy researchers RM Forrest & AL Greer have applied the fourth paradigm of scientific research by training a machine-learning model of glass-forming ability on a large dataset of experimental measurements, not merely to produce a predictive tool, but also to extract theoretical insights. While recovering several established empirical criteria, they identified an important relation between glass-forming ability and the electronic structures of the elements in an alloy. Novel alloy compositions with high glass-forming ability and also crystalline high-entropy alloys may be systematically designed by selecting elements to induce mutual changes in atomic radii.Metallic glasses are created from molten metal alloys by cooling at a rate fast enough to prevent any significant degree of crystallization. The lack of time for transition into a crystalline ordering leaves the atoms in a liquid-like structure.2 Alloy compositions for which it is possible to produce samples with a minimum thickness of 1–10 millimeters are referred to as bulk metallic glasses, while others may be obtainable only as thin glassy ribbons. The amorphous structures of metallic glasses give them interesting properties, with potential applications in many areas, from sporting equipment to aircraft and automotive components. The relative novelty of metallic glasses s means further glass-forming alloy compositions are awaiting discovery, with the promise of meeting previously inaccessible requirements for applications.The ability of an alloy composition to resist arrangement into an ordered crystalline phase upon cooling is referred to as the glass-forming ability. The glass-forming ability of an alloy composition may be directly quantified via the maximum achievable diameter of a rod cast into a fully glassy state, Dmax. Predicting the glass-forming ability of alloy compositions is a central goal in MG research. This research is plagued by empirical rules and trial-and-error experimentation, with frustratingly many proposed criteria for glass-forming ability, each often published with claims of superiority yet limited proven applicability. The present work explores the power of the burgeoning ‘fourth paradigm’ of scientific discovery, that being the utilization of data to train machine-learning models of physical phenomena, which in turn inform our theoretical understanding.Machine-learning involves the creation of models that can improve their performance via exposure to data. Materials science is no exception to the rapid spread of Machine-learning as a research tool. Machine-learning approaches to materials science use the vast amount of experimental data now available to model the physical laws governing observed phenomena. Neural networks, in particular, are commonly applied in Machine-learning and have been widely used within materials science to model atomic interactions, predict synthesis routes,15 reconstruct structures from imaging, identify phases and transitions, predict welding criteria, predict material properties and to predict the existence of novel materials, among others.Researchers said “Leveraging this insight and the predictive ability of the ensemble neural-network model, we make a number of suggestions for novel binary and ternary glass-forming alloy systems that will exhibit increases in the deviation of atomic radii due to the balancing of electron density. In the search for these alloy systems, lithium frequently appears due its combination of small radius and low nWS, and has been seen elsewhere in the literature as an addition to alloy compositions which increases glass-forming ability. We emphasize the need for more data to be published on glass-forming alloy compositions, both detailing the successful creation of glassy alloys but, importantly, also failed attempts. This is essential to avoid the study of BMGs becoming entrenched in well-trodden areas of composition-space, dependent on empirical rules. Further, the relative lack of BMG-forming alloy compositions with large Dmax presents a significant challenge to ML, as little information is provided from which to learn the rules that determine their existence. Nevertheless, the transferability of our model is successfully tested via k-folds cross-validation. Future work in this direction may consider more advanced theories than the Thomas–Fermi model for electron density, such as the Thomas–Fermi–Dirac model which includes the exchange energy, or a full density-functional theory treatment. In addition to nWS, we identify several other features to be important to the neural-network model, of which deeper investigation may return further useful insights into glass formation.”
