Artificial Intelligence helps identify bulk metallic glasses with exceptional hardness
Journal Nature Communications selects Max Planck study for its Editors’ Highlights collection
At a glance:
- Award: Journal Nature Communications has selected the study by Anurag Bajpai and colleagues for the journal’s Editors’ Highlights collection.
- Research topic: How can researchers design bulk metallic glasses with exceptional hardness when experiments are costly and high-quality data are scarce?
- Approach: New artificial intelligence approach combines attention-based learning and uncertainty-aware inverse alloy design to navigate high-dimensional alloy-composition spaces while screening candidates for chemical plausibility, novelty and predictive reliability for exceptional hardness. .
- Result: Newly designed bulk metallic glasses achieved around 2,450 HV in Vickers hardness, twice the hardness of many high-hardness Fe- and Co-based bulk metallic glasses and substantially above typical Zr-based bulk metallic glasses. This places it among the hardest bulk metallic glasses reported under comparable testing conditions. Dense atomic packing, boron-rich local environments and refractory elements such as niobium and tungsten contribute to the exceptional hardness.
- Outlook: The approach could help accelerate the discovery of high-performance materials when experimental data is scarce.
How can artificial intelligence move beyond predicting known alloys and propose new compositions that can actually be made? Anurag Bajpai and his colleagues at the Max Planck Institute for Sustainable Materials (MPI-SusMat) have developed an artificial intelligence (AI) approach for the inverse design of exceptionally hard bulk metallic glasses. The framework learned a chemically structured design landscape and searched it for candidates that balanced hardness, novelty, chemical plausibility and predictive uncertainty. Bulk Metallic glasses can be used for wear-resistant coatings, precision mechanical components and microelectromechanical systems. Their results were published in the journal Nature Communications and were selected for the journal’s Editors’ Highlights collection.
An AI framework that learns what matters and how certain it is
Metallic glasses are alloys in which the atoms are frozen into a disordered arrangement rather than a periodic crystal lattice. They combine high hardness with elastic resilience and corrosion resistance. Particularly hardness indicates resistance to permanent deformation and surface damage.
In crystalline alloys, hardness can often be related to dislocations, grain boundaries and microstructural characteristics. Metallic glasses contain none of these conventional features. Instead, plastic flow begins through local atomic rearrangements known as shear-transformation zones, whose activation depends on short-range packing, bonding and excess free volume. Changes of only a few atomic percent can therefore produce disproportionate changes in hardness, while the number of possible compositions grows as more elements are added.
“Instead of testing thousands of possible compositions experimentally, we wanted to develop a method that could intelligently navigate this vast design space and identify the most promising candidates,” explains Anurag Bajpai, group leader at MPI-SusMat and first author of the highlighted publication. The critical point is to search for novelty in new alloys without leaving the region where the chemistry remains plausible and the model’s uncertainty remains acceptable.
The new AI framework first uses an attention mechanism to learn which elemental contributions are most relevant to hardness within the available data. A variational information bottleneck then compresses the alloy composition and indentation load into a low-dimensional representation that retains the hardness-relevant signal while suppressing redundant or noise-dominated patterns. Rather than generating unrestricted compositions, the framework searches this learned space under explicit constraints on chemical plausibility, novelty and uncertainty. This allows the researchers to distinguish between candidates that the model predicts with confidence and those for which it has little evidence.
“This is particularly important when searching for materials beyond the range of the existing data,” says Bajpai. “The model should not simply be confident because a composition looks promising mathematically. It also needs to indicate how chemically credible that prediction is and when the available evidence is weak.”
Bulk metallic glasses with exceptional hardness
Using this approach, the researchers designed and experimentally produced five new multicomponent metallic glasses that were rich in boron, niobium, iron, tungsten and one of several refractory elements, including cobalt, hafnium, ruthenium or zirconium. All five form fully amorphous rods with a diameter of 2 mm and reach Vickers hardness values of around 2,450 HV, which are among the highest reported for bulk metallic glasses and nearly twice the 11–13 GPa commonly reported for many Fe- and Co-based bulk metallic glasses, under comparable testing conditions. To understand the physical origin of the hardness, the team combined high-energy X-ray scattering with molecular-dynamics simulations. The hardness correlated with tighter first-neighbour packing, boron-enriched short-range environments, and local rigidity stabilized by refractory and transition-metal additions: features found to be consistent with greater resistance to the atomic rearrangements that initiate plastic flow.
The work, therefore, moves beyond a conventional “predict and test” approach towards AI-guided inverse materials design, starting with a desired property and working backward to identify compositions that could deliver it. The developed AI framework is not only applicable to hardness in metallic glasses but can be used for any alloy class and target property.
