Sebastian Alexander Kube

Predicting metallic glass-forming ability (GFA) through ML has been challenging. Among the reasons for the challenge are limited, biased and inconsistent experimental data, binary vs continuous labeling, and limited feature representation particularly for multicomponent alloys. To address these challenges, we restrict the problem to binary alloys and base the training data on experiments spanning a representative fraction of ~ 20% of binary alloys, which have been synthesized through sputtering and its atomic structure characterized. Consistent in quality and sufficient quantity of binary sputtering data with continuously labeled GFA quantified in the critical cooling rates (Rc) are constructed, and Physically Informed features are employed to describe the alloy in terms of its Rc. Using the PI features, the RF model predicts Rc with MAPE of ~18% (log₁₀ (Rc) space), compared to a benchmark model (28%). However, attempts to use this binary model to predict Rc in ternary alloys are ineffective; using the ML model trained on binary data to predict Rc in ternary alloys results in a MAPE of ~36% for sputtered ternary alloys and ~ 89% for ternary BMG formers. Such poor performance suggests that the mixing complexity in ternary alloys is not captured in binary alloys. Overall, while addressing previous challenges in predicting GFA through ML, the models achieve statistically improved predictive performance. However, the achieved predictions remain coarse, indicating additional limitations that constrain precise Rc prediction. Most fundamentally, the discreteness of atoms remains the limiting factor from using ML strategies to precisely predict glass-forming ability of alloys.

Developing new high-performance alloy compositions is expensive and time-consuming. Conventional experimental alloy testing performed manually by human researchers is a primary bottleneck. To overcome this, we are building “AlloyBot”, a self-driving laboratory that can autonomously synthesize and test alloys 100× faster. Here we present the first building block: a fully automatic weighing and arc-melting system that can synthesize 1 alloy per hour (>150 alloys per week). Various innovations facilitate this process, such as a load-lock system to rapidly exchange crucibles, and four electrodes used for arc-melting. Alloys can be synthesized as buttons or cast rods, up to 25cm ³ in volume (>100g Ti-6Al-4V), and at a unit cost broadly ranging around ~100 USD/5cm ³ , driven primarily by feedstock pricing rather than human labor. Within metallurgical constraints, any element combination and composition can be synthesized on-demand. Performance is validated through the synthesis of Ti-6Al-4V alloy, Ti ₂₀ Zr ₂₀ Hf ₂₀ Nb ₂₀ Ta ₂₀ refractory high-entropy alloy, and Zr ₅₅ Cu ₃₀ Ni ₅ Al ₁₀ metallic glass.

Tailoring material properties often requires understanding the solidification process. Herein, we introduce the geometric descriptor Soliquidy, which numerically captures the Euclidean transport cost between the translationally disordered versus ordered states of a materials. As a testbed, we apply Soliquidy to the classification of glass-forming metal alloys. By extending and combining an experimental library of metallic thin-films (glass/no-glass) with the aflow.org computational database (geometrical and energetic information of mixtures) we found that the combination of Soliquity and formation enthalpies generates an effective classifier for glass formation. Such classifier is then used to tackle a public dataset of metallic glasses showing that the glass-agnostic assumptions of Soliquity can be useful for understanding kinetically-controlled phase transitions.

The underlying atomistic mechanism of deformation is a central problem in mechanics and materials science. Whereas deformation of crystalline metals is fundamentally understood, the understanding of deformation of amorphous metals lacks behind, particularly identifying the involved temporal and spatial scales. Here, we reveal that at small scales the size-dependent deformation behavior of amorphous metals significantly deviates from homogeneous flow, exhibiting increasing deformation rate with reducing size and gradually shifted composition. This transition suggests the deformation mechanism changes from collective atomic transport by viscous flow to individual atomic transport through interface diffusion. The critical length scale of the transition is temperature dependent, exhibiting a maximum at the glass transition. While viscous flow does not discriminate among alloy constituents, diffusion does and the constituent element with higher diffusivity deforms faster. Our findings yield insights into nano-mechanics and glass physics and may suggest alternative processing methods to epitaxially grow metallic glasses.