Machine learning and 3D printing forge titanium stronger than any before it

Engineers have long dreamed of metals whose internal architecture is engineered at the nanometre scale, where dense arrays of boundaries and defects can push strength far beyond what conventional metallurgy allows. A team led by researchers at City University of Hong Kong has now brought that dream closer to industrial reality, reporting in Nature Materials a titanium alloy produced by laser powder bed fusion that reaches a yield strength of roughly 1.5 gigapascals and an ultimate tensile strength exceeding 1.7 gigapascals, while still stretching uniformly by about 7.5 percent before necking. The secret lies in a previously unseen microstructure: parallel bands of alternating crystal stacking sequences, each thinner than ten nanometres, woven through every martensitic lath of the as-built part.

The achievement matters because nanostructured metals have always faced a manufacturing bottleneck. Laboratory routes such as sputtering, electrodeposition and severe plastic deformation can create spectacular nanoscale architectures, but they are constrained by geometry and low throughput, leaving a persistent gap between proof-of-concept demonstrations and real components. Forming a nanostructure demands extreme, non-equilibrium conditions, such as intense temperature, stress or compositional fields that drive crystalline defects to extraordinary densities, and industrial processes inevitably steer materials back toward thermodynamic equilibrium. Bulk components with complex shapes that retain their nanostructure have therefore remained elusive.

Laser powder bed fusion, a form of metal additive manufacturing, offers a way out. By selectively melting powder layer by layer under programmed scanning strategies, the process imposes extreme thermal gradients, rapid cooling rates and repeated thermal cycling, conditions far from equilibrium in which ultrafine grains, metastable phases and nanoscale features can be stabilized inside bulk parts. Additive manufacturing also sidesteps the segregation and formability limits of ingot casting, opening compositional windows that conventional processing cannot reach. Yet past efforts largely adapted existing alloy chemistries, and the resulting features typically remained coarser than the ideal nanoscale regime where strengthening is most efficient.

To navigate the immense search space of compositions and process parameters, the team combined high-throughput experimentation with machine learning. They chose the Ti–Al–Cr system, using aluminium as an alpha-phase stabilizer and chromium as a beta-phase stabilizer, and blended elemental powders of commercially pure titanium, aluminium and chromium for in situ alloying during printing. Chromium is particularly intriguing because it stabilizes the beta phase effectively but has been limited in wrought or cast titanium alloys by segregation and processing incompatibility, constraints that additive manufacturing lifts. Laser power and scanning speed were identified as the most influential process variables and incorporated alongside composition as primary design inputs.

In total, 118 distinct Ti–Al–Cr alloys were fabricated by varying aluminium and chromium contents together with laser power and scan speed, each tested under uniform protocols to produce a consistent dataset. Machine learning models were trained with compositions and processing parameters as inputs and ultimate tensile strength, yield strength and uniform elongation as outputs, enriched with two empirical metallurgical descriptors: molybdenum equivalence for beta-phase stability and a solid-solution-strengthening parameter. Among five algorithms evaluated by leave-one-out cross-validation, XGBoost performed best, achieving a root mean square error within 50 megapascals and coefficients of determination of 0.87 for strength and 0.97 for yield strength, with most predictions within five percent of experimental values.

The trained model then screened 38,052 virtual alloy-process combinations spanning the design space. Candidates in the top five percent of predicted strength were validated experimentally, and all exceeded 1,630 megapascals in tensile strength, well above previously reported additively manufactured titanium alloys. Shapley additive explanation analysis revealed chromium content as the dominant feature governing strength under laser powder bed fusion conditions. Notably, the winning compositions fall outside the conventional titanium alloy families historically optimized for casting or wrought processing, underscoring that alloys tailored to additive manufacturing require design strategies beyond established practice.

The star performer, a Ti–7.5Al–3.3Cr alloy printed at 120 watts and 1,750 millimetres per second, owes its properties to a hierarchical nanomartensitic architecture. Electron backscatter diffraction showed the microstructure is 99.1 percent alpha-prime martensite with a negligible beta fraction, and scanning transmission electron microscopy revealed that within laths roughly 75 nanometres thick lies an unexpected internal substructure: parallel nano-bands thinner than ten nanometres, oriented normal to the hexagonal c-axis. Atomic-resolution imaging showed these bands comprise densely arranged AB-, BC- and CA-type stacking variants of the hexagonal close-packed lattice, separated by boundaries where single-period face-centred-cubic stacking units necessarily appear when one sequence transforms into another. Atom probe tomography confirmed lamellar concentration modulations of aluminium and chromium with a spacing of about six nanometres, matching the band periodicity.

Molecular dynamics simulations explained the origin. Under equilibrium annealing, diffusional transformation produces coarse, defect-free hexagonal laths, but rapid cooling drives a martensitic pathway in which large undercooling generates densely packed laths full of stacking faults. Multiple martensite variants nucleate from the body-centred-cubic parent phase, and rapid expansion from multiple nucleation sites within each lath creates distinct stacking sequences whose intergrowth produces the hetero-stacking bands. The team validated this kinetic origin by reheating the alloy and applying different cooling rates: slow cooling yielded near-equilibrium structures, faster cooling produced martensite, and only the extreme quench rates of laser powder bed fusion fully developed the sub-10-nanometre architecture.

Those bands are what make the alloy extraordinary. In simulations, dislocations crossing the hetero-stacking boundaries must dissociate into partials, fragmenting into sub-10-nanometre segments that are pinned at the interfaces, while the bands simultaneously act as dislocation sources, promoting multiplication and multiple slip systems that evolve into entangled networks. This delocalizes plastic strain and sustains work hardening, whereas conventional titanium shows facile planar slip with severe strain localization. Post-deformation microscopy confirmed the picture: after 7.5 percent strain the alloy contained a high density of a-type dislocations pinned at band boundaries and long, entangled a-plus-c-type dislocations, with a-plus-c slip bands even transmitting across adjacent laths and restoring uniform stacking as they passed. In situ testing showed shear bands spreading uniformly into a dense interwoven network, and fracture surfaces displayed evenly distributed dimples.

The result is a directly printed titanium alloy with a specific strength of about 380 megapascals per cubic centimetre per gram, far above additively manufactured steels, aluminium alloys, nickel alloys and high-entropy alloys, and tensile strength above 1,100 megapascals retained up to 550 degrees Celsius, extending the performance envelope of printed titanium into elevated-temperature service. Compositional tuning through in situ alloying also offers tailored trade-offs: Ti–7.5Al–5Cr reaches 1,555 megapascals with 12.4 percent elongation, and Ti–7.5Al–7.5Cr achieves 1,449 megapascals with 15.9 percent elongation through transformation-induced plasticity. Beyond the alloy itself, the fabricate–predict–optimize cycle demonstrated here turns alloy development from empirical trial and error into a predictive, data-driven workflow, offering a transformative pathway for discovering next-generation metals that combine nanoscale architecture with scalable manufacturing.

Subject of Research: Machine learning-guided additive manufacturing of ultrastrong nanostructured titanium alloys

Article Title: Sub-10-nm stacking bands enable ultrastrong additively manufactured titanium

Article References: Bai, X., Hua, D., Luan, H., Chen, X., Duan, F., Xing, H., Zhao, D., Xu, Z., Li, S., Li, Y., Luan, J., He, C., & Lu, J. (2026). Sub-10-nm stacking bands enable ultrastrong additively manufactured titanium. Nature Materials. https://doi.org/10.1038/s41563-026-02749-6

Image Credits: AI Generated

DOI: 10.1038/s41563-026-02749-6

Keywords: titanium alloys, additive manufacturing, laser powder bed fusion, machine learning, nanostructure, martensite, stacking faults, high strength, work hardening, alloy design, high-throughput experimentation, materials science