Repository logo

On the application of machine-learning to the study of glass-formation in metallic systems: Prediction, optimisation, and characterisation of novel glassy alloys



Change log



Metallic glasses (MGs) are amorphous materials created by quenching a molten alloy mixture so quickly that crystalline phases do not have time to nucleate and grow, leaving a solid material with liquid-like structural disorder. This lack of crystalline ordering in MGs gives rise to a variety of desirable properties, including strong corrosion resistance and excellent soft magnetism, and leads to diverse potential applications, from golf clubs to transformer cores.

The proclivity of any particular alloy composition to form a glass is known as the glass-forming ability (GFA). This is, however, poorly understood, greatly restricting the development of novel alloys. Many empirical rules and criteria aiming to describe the GFA of alloy compositions have emerged since MGs were first discovered in 1959, but they have limited predictive power and do little to demonstrate fundamental understanding.

This thesis explores the application of the ‘fourth paradigm’ of scientific research, that being the use of machine-learning (ML) to model relationships within large datasets, to the study of glass formation in metallic systems, with the dual aims of improving understanding of the phenomenon of glass formation itself, and accelerating the identification of novel glass-forming alloys.

In this thesis, a neural-network model is applied to predict simultaneously the liquidus temperature, glass-transition temperature, crystallisation-onset temperature, maximum glassy casting diameter, and probability of glass formation, for any given alloy. Feature permutation is used to identify the features of importance in the black-box neural network, recovering Inoue’s empirical rules, and highlighting the effect of discontinuous Wigner-Seitz boundary electron-densities on atomic radii. With certain combinations of elements, atomic radii of different species contract and expand to balance electron-density discontinuities such that the overall difference in atomic radii increases, improving GFA. Adjusted atomic radii are calculated via the Thomas-Fermi model and this insight is used to propose promising novel glass-forming alloy systems.

Building on the predictive capabilities enabled by the ensemble neural-network model, further new glass-forming alloys are identified by means of a genetic algorithm (GA); the genetic operators of competition, recombination, and mutation to a population of trial alloy compositions, with the goal of evolving towards candidates with excellent glass-forming ability. The size of composition space means even coarse grid-based searches for interesting alloys are infeasible unless heavily constrained, which requires prior knowledge and reduces the probability of making novel discoveries. GAs provide a practical alternative to brute-force searching, by rapidly homing in on fruitful regions and discarding others. In this work, optimisation focuses on the maximum casting diameter of a fully glassy rod, the width of the supercooled region, and the price-per-kilogramme, to identify commercially viable novel glass-formers. The GA is also applied with specific constraints, to identify aluminium-based and copper–zirconium-based glass-forming alloys, and to optimise existing zirconium-based alloys. In particular, aluminium-lanthanide alloys are identified to be of relatively low cost-per-kilogramme and high GFA.

Finally, this thesis studies the properties of crystalline and glassy alloys via atomistic simulations. Ab initio quantum-mechanical calculations are typically too computationally expensive for investigations of glassy materials, requiring the use of approximate inter-atomic potentials. Due to the novelty of the alloy composition selected from GA output, namely aluminium-ytterbium, the creation of new interatomic potentials is required. Classical embedded-atom-method (EAM) potentials and machine-learning-based Gaussian approximation potentials (GAPs) and spectral-neighbour-analysis potentials (SNAPs) are generated, and used to model the material properties of Al-Yb in crystalline and glassy forms, producing predictions including equilibrium densities and elastic moduli. Characterisation of the properties of these materials enables the suggestion of engineering applications for which they would be suitable, with the glasses predicted to be competitive with crystalline materials for a variety of uses.





Greer, A Lindsay


Interatomic Potentials, Machine learning, Materials Discovery, Metallic glasses


Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge
European Research Council (695487)