Researchers Set the Alloy System, Active Learning Picked the Compositions: HEA05 Reaches 1.75 GPa Yield Strength with 25% Uniform Elongation

In June 2025, a Xi'an Jiaotong University team reported the multi-principal-element alloy Fe₃₅Ni₂₉Co₂₁Al₁₂Ta₃, called HEA05, in Nature. The team fabricated, rolled and heat-treated the alloy, then performed room-temperature tensile testing and multiscale microstructural characterization.^[S1]

After a representative one-hour ageing treatment at 750 °C, HEA05 reached a yield strength of 1,750 ± 50 MPa, an ultimate tensile strength of 2,403 ± 46 MPa and a uniform elongation of 25 ± 1.5%. True stress approached 3 GPa before fracture, while the work-hardening rate remained above 2 GPa across a large strain range.^[S1]

Bulk metals near 2 GPa rarely retain large uniform plasticity. In some ultra-high-strength steels, nominal elongation includes localized deformation from Lüders or Portevin–Le Châtelier bands, while conventional precipitation-strengthened high- and medium-entropy alloys usually remain below roughly 1.2 GPa in yield strength. HEA05 deforms smoothly, and the same composition can be processed to span combinations from 1.5 GPa with 31% uniform elongation to 1.95 GPa with 15%.^[S1]

66.6% + 15%

volume fractions of L1₂ and B2^[S1]

1.5 / 31% → 1.95 / 15%

strength (GPa) / uniform-elongation range from one composition under different processing^[S1]

The model did not invent an alloy from the entire periodic table

The paper estimates a potential multi-component alloy space of roughly 10⁴⁹ variants; that is a broad count of possible combinations, not the space the model actually searched. Instead of handing that space to a black box, the team selected an FeNiCo FCC matrix and added Al and Ta, aiming for solid-solution strengthening, a high volume fraction of coherent L1₂ nanoprecipitates and a controlled amount of deformable B2 phase.^[S1]

Active learning began with literature data on 140 FCC AlCoCrFeNiTa high-entropy alloys. The researchers confined each element to a set range (Al 5–13 at.%, Ta 1–3 at.%, for example), retained six of twenty candidate physical features and filtered virtual compositions with 7.8 ≤ VEC ≤ 8.4; after each round they also adjusted the principal elements and composition ranges. A random-forest surrogate proposed two candidates per round, one purely on the highest prediction and one that also weighed uncertainty. The researchers melted, cold-rolled, recrystallized, aged and tensile-tested each before the measurements returned to the training set. Across four rounds, eight new compositions were fabricated and tested, the first round's two reached 900 and 685 MPa in yield strength and the later six 1,288–1,750 MPa, and HEA05 appeared in round three. Repeat tensile tests, alternative heat treatments and microstructural characterization were further experiments on top of those eight compositions.^[S1][S2]

The supplementary information also records an earlier attempt that failed: four rounds in the Ti-containing AlCoCrFeNiTi system (200 literature alloys) produced nothing that beat the training data. The researchers judged Ti a weaker L1₂ strengthener than Ta, swapped Ta in and retrained; only then did the model point them to high-Al compositions and reach the strength goal in round three. Which element to use was a human decision.^[S1]

That boundary matters. The model primarily searched chemical composition with an emphasis on yield strength. It did not directly predict “1.75 GPa plus 25%,” nor did it autonomously produce the final three-phase microstructure. Rolling, recrystallization and ageing were still designed by researchers. The more accurate sequence is: domain knowledge narrows the space, the model selects compositions, experiments return data, and microstructure engineering produces the final properties.^[S1]

The L1₂ + B2 dual-precipitate structure carries the performance

Representative HEA05 contains an FCC matrix with an average grain size of about 7.6 μm and two precipitate populations: coherent L1₂ nanoprecipitates roughly 15 ± 3 nm in size at 66.6 ± 2.3% volume fraction, and incoherent B2 precipitates roughly 350 ± 70 nm at 15 ± 2%.^[S1]

The authors estimate that L1₂ contributes about 859 MPa to yield strength and B2 about 351 MPa, with the remainder coming from lattice friction, pre-existing dislocations and grain boundaries. The high-volume-fraction L1₂ phase is the primary strengthening source, while Ta raises its antiphase-boundary energy and makes ordered precipitates harder for dislocations to shear.^[S1]

B2 behaves less conventionally. Traditional NiAl-type B2 is hard and often brittle. The multi-component B2 in HEA05 has lower chemical ordering energy, with Ta and other alloying elements reducing antiphase-boundary energy by about 30% relative to binary NiAl. Dislocations can therefore enter B2, multiply and be stored during tension. B2 raises yield strength and also supports later work hardening.^[S1]

Why 25% uniform elongation is the result to notice

For uniform tensile deformation to continue, work hardening must offset the instability caused by a shrinking cross-section. HEA05 hardens from 1.75 GPa yield strength to 2.40 GPa ultimate strength, a gap of roughly 650 MPa and a yield ratio of about 0.73. Its work-hardening curve shows no pronounced serrations or collapse.^[S1]

Geometrically necessary dislocation density in B2 rises from about 0.8 × 10¹⁴ m⁻² before deformation to 6.8 × 10¹⁴ m⁻² at 20% strain, even above the roughly 5.4 × 10¹⁴ m⁻² measured in FCC/L1₂ regions at the same strain. Strain incompatibility among phases also produces hetero-deformation-induced hardening. The alloy is not merely strong at yield; it continues storing dislocations throughout plastic deformation.^[S1]

This is small-data AI4S. A set of 140 samples is tiny by modern machine-learning standards, but complete composition–processing–microstructure–property labels are expensive in materials science. Instead of masking data scarcity with a more elaborate model, the team used knowledge of VEC, atomic-size mismatch, phase stability and precipitation strengthening to prune the search space.^[S1]

When a model can generate many candidates but the laboratory can melt and test only two per round, generating more compositions is no longer the tightest constraint. Active learning becomes a way to allocate experimental budget: choose the candidate most likely to improve performance, or the one whose failure would still reduce uncertainty most effectively.

HEA05 is not a self-driving laboratory; melting, processing, characterization and judgment still rely heavily on people. But it closes an important part of the loop: the model proposes compositions, physical experiments return results, and the next round changes accordingly. Metrics such as experiments-to-discovery, time-to-verification and information gained per experiment may matter more here than R² or RMSE alone.

This is not an AI-designed material supported only by computational metrics.

HEA05 was physically fabricated and repeatedly tested in room-temperature tension (at least three samples per condition; the paper does not say whether ± is a standard deviation or another spread). Its microstructure was characterized with XRD, TEM, SEM, EBSD, TKD, APT and synchrotron scattering. The 1.75 GPa and 25% values are measurements, not model outputs, and the training data and code are public.^[S1][S2]

Machine-learning design should be interpreted narrowly.

The model did not directly optimize both 1.75 GPa and 25%, nor did it predict the final three-phase microstructure end to end. Experts chose the element system, physical features and VEC constraints, while later processing determined the final structure. This is ML-assisted alloy design, not fully autonomous discovery.^[S1]

Validation still stops early in the path to an engineering material.

The paper validates room-temperature strength–ductility and deformation mechanisms, but not long-term high-temperature stability, fatigue life, fracture toughness, corrosion resistance, manufacturing at scale or industrial qualification. HEA05 also contains 3 at.% Ta. Reducing cost and reproducing the same precipitate architecture in large products remain open problems.^[S1]

Domain knowledge can become part of the search algorithm

Materials knowledge defines the element system, physical features and phase constraints; machine learning chooses the next expensive experiment within the remaining space. In small-data science, that division is more practical than treating domain knowledge and models as opposites.

Active learning allocates scarce physical labels

Model inference is cheap; melting, heat treatment, characterization and tensile testing are expensive. Active learning asks a concrete question: if only two experiments fit in this round, what is most worth asking nature next?

Finding a composition and explaining its behavior are different tasks

Active learning found the HEA05 composition, but later experiments explained how L1₂ strengthens it, why B2 does not simply embrittle it and how B2 stores dislocations at later strain. Prediction and scientific understanding still require separate evidence in AI4S.

A hard second phase does not have to act only as an obstacle

By lowering B2 ordering energy through multi-component alloying, an intermetallic phase usually considered hard and brittle can deform, store dislocations and sustain work hardening. That materials insight is more memorable than the phrase “AI found a record-setting alloy.”