Bio-inspired hierarchical microstructures offer a route towards engineered fatigue resistance in additively manufactured alloys. However, it remains unclear how discrete structural constituents independently govern damage accumulation, particularly during the critical fatigue initiation regime where short cracks strongly interact with local microstructure. Here, we investigate multiscale fatigue initiation in a dual-phase, nanolamellar AlCoCrFeNi_2.1 high-entropy alloy. By comparing microscale specimens that isolate the nanolamellar structure against macroscale specimens containing the full melt-pool architecture, we identify size-dependent fatigue initiation mechanisms. We find that failure is dictated by nanolamellar interfaces at the microscale, whereas mesoscale melt pool boundaries serve to initiate fatigue at the macroscale. This mechanistic shift is accompanied by a transition from macroscale quasi-brittle failure to microscale plasticity-driven crack extension. Our results provide a physical framework for understanding how structural hierarchy governs the transition from discrete microstructural deformation to continuum fatigue fracture behavior, informing the design of damage-tolerant, additively manufactured alloys.

Materials manufacturing

AlCoCrFeNi_2.1 eutectic high-entropy alloy (HEA) samples were manufactured using an EOS M290 Laser Powder Bed Fusion (L-PBF) system (EOS GmbH, Germany) equipped with a 400 W Yb-fiber laser. The laser beam was focused through an F-theta lens, producing a fixed spot size of 100 μm. Gas-atomized pre-alloyed powders with particle sizes ranging from 15 to 53 μm were employed as the feedstock. All samples were processed under an argon atmosphere with an oxygen concentration maintained below 1000 ppm. A 4140 steel substrate was preheated to 80 °C to mitigate thermal residual stresses and enhance interfacial bonding. The L-PBF process was conducted using a laser power of 350 W, scanning speed of 1000 mm/s, hatch spacing of 0.08 mm, and layer thickness of 0.04 mm. A continuous stripe scanning strategy with full-length laser tracks across the build area was adopted, together with a 67° interlayer rotation, to minimize residual stress accumulation and promote microstructural uniformity. These optimized printing parameters match those reported in prior work. Rectangular plates with dimensions of 50 × 20 × 3 mm³ were fabricated, achieving a relative density exceeding 99.5 %.

As-cast samples were prepared by arc-melting a mixture of raw elements with a purity of approximately 99.9 wt. % under an argon atmosphere. The ingots were inverted and re-melted at least six times to ensure compositional homogeneity, followed by suction casting into a water-cooled copper mold with dimensions of 50 × 10 × 3 mm³.

Specimen fabrication

Two perpendicular edges of bulk AM and cast samples were manually polished with SiC sandpaper (800, 1200 grit), followed by a water-based diamond suspension (9 μm, 1 μm) and colloidal silica (0.05 μm). Microcantilevers were subsequently fabricated using a Tescan S8252X FIB-SEM (TESCAN, Czech Republic) with a Xe⁺ ion plasma source at a constant accelerating voltage of 30 keV. Rough cuts were performed at 100 nA, 30 nA, and 10 nA, followed by polishing cuts at 1 nA and 100 pA on all faces. A notch was introduced with a top-down, 100 pA cut in the final step. For AM specimens, notches were positioned near (but not intersecting) melt pool boundaries to achieve a predominantly orthogonal lamellar orientation relative to the crack front. Following fabrication, imaging was performed using the electron beam of the same instrument with a constant accelerating voltage of 5 keV, 300 pA beam current to extract the specimen dimensions. Additional high resolution imaging was performed using a Zeiss ORION NanoFab Scanning Helium Ion Microscope (ZEISS, Germany) at 25 kV accelerating voltage. Further post-mortem characterization was performed using a Keyence VK-X3000 Series 3D Surface Profiler (Keyence, USA).

Single-edge notched bend (SENB) specimens were machined exclusively from the as-built AM samples using wire EDM. Specimens had a side-to-side thickness of B = 2.5 mm oriented along the build direction, width W = 5.0 mm, length L = 22.5 mm, and an initial crack length a₀ = 2.5 mm, in compliance with ASTM E1820 guidelines. Crack propagation occurred along W, orthogonal to the build direction.

Nanoindentation experiments

Nanoindentation experiments were performed on bulk, polished AM and cast specimens using an iMicro Nanoindenter (KLA, USA) with an InForce 50 actuator and a 1.89 μm radius diamond conical tip (Synton-MDP, Switzerland). All tests were performed at a constant strain rate of 1 × 10^-3 s^-1 to a peak load of 8 mN, using a 100 Hz continuous stiffness measurement (CSM) signal at a target displacement amplitude of 2 nm to monitor stiffness. Using the Oliver and Pharr method, elastic moduli of E = 142.8 ± 13.6 GPa (n = 32 measurements) and E = 50.3 ± 2.8 GPa (n = 24) were measured for the AM and cast systems, respectively. Indentation depth, h_indent, and contact stiffness, S_contact, were expressed as functions of the indentation load, P, by fitting power-law equations to averaged nanoindentation data.

Microscale fatigue experiments

Microcantilever fatigue experiments were performed using a NanoFlip Nanoindenter (KLA, USA) with an InForce 50 actuator and the same conical tip used for nanoindentation, integrated with a Quanta 600 FEG Mark II ESEM (FEI Company, USA). The SEM stage was tilted by 5-10° relative to the electron beam to improve visibility. Videos were collected at 10 FPS with a beam dwell time of either 300 ns/pixel or 1 μs/pixel at 10 keV accelerating voltage. Each microcantilever was loaded monotonically to a target static load, P_static, at a constant displacement rate of 10 nm/s, which was allowed to decay at 0.1 mN/μm to maintain stable actuator control during dynamic loading. A 10 Hz sinusoidal load, P_dynamic, with RMS amplitude of 2 mN was superimposed on the static load to induce fatigue and monitor stiffness, leading to a varying stress ratio. Additional experiments (AM4, C4) were performed using a 1 Hz cyclic loading method with a constant stress ratio, R = 0.1; for these tests, stiffness was measured from the unloading slope of each cycle.

Stiffness data were corrected by treating machine compliance, contact stiffness, and beam bending stiffness as a system of springs in series. Machine compliance was determined from instrument calibration, while the contact stiffness was measured from nanoindentation data. The first cycle (N = 1) was defined at the first measurable decrease in stiffness from the initial value. Crack length, a, was computed from the corrected bending stiffness using elastic-plastic finite element (FE) simulations relating normalized stiffness S/S₀ to normalized crack length a/W. The resulting a vs. N data were smoothed using a fourth-order univariate spline. Crack growth rates, da/dN, were computed as the analytical derivative of this spline. The driving force was estimated from the cyclic ranges of load, ΔP, and displacement, Δd, for each cycle. These ranges were substituted into the area-based definitions for J standardized in ASTM E1820.

Macroscale fatigue experiments

SENB fatigue experiments were performed using a MTS Criterion Model 43 uniaxial testing machine (MTS Systems Corporation, USA) equipped with a 10 kN load cell. Specimens were tested with a custom-fabricated steel three-point bend fixture with a support span, S = 20 mm. Both the support and loading pins were stainless steel cylinders with a nominal diameter of 1.5 mm. Cyclic loading was performed with a stress ratio, R = 0.1, and a peak load, P_max = 521 N, determined using the ASTM E1820 fatigue pre-crack load formula. ASTM E1820 permits fatigue loading at 0.7P_max to prevent premature failure; specimen B1 was tested for 4000 cycles at this reduced load with no crack extension, so all subsequent testing was performed at P_max. All tests were performed at a displacement rate of ±5 mm/minute, corresponding to a cycle frequency of approximately 0.4 Hz. The crack tip was imaged during testing using a BFS-U3-88S6M-C USB 3.1 Blackfly S camera (Teledyne FLIR, USA) with an adjustable zoom lens, recording at a frame rate of 14.5 Hz with a resolution of 1920 × 1080 pixels.

The load-line compliance, C, was obtained from the inverse slope of the unloading data for each cycle and corrected for machine compliance. Normalized crack length, a/W, was computed from the corrected compliance following ASTM E1820 and validated against crack mouth opening compliance measurements. The first cycle (N = 1) was defined at the first measurable increase in compliance from the initial value. Both crack length, a, and load, P vs. N data were smoothed using a Gaussian filter with σ = 100 cycles. Crack growth rates, da/dN, were computed as the numerical derivative of the smoothed crack length data using the ASTM E647 Incremental Polynomial Method with n = 4. The change in linear elastic stress intensity factor during cycling, ΔK = K_max - K_min, was computed using the smoothed data from the ASTM E1820 expression for SENB specimens, with the load range taken as 0.9 P_max for R = 0.1. The provisional fracture toughness, K_Q, was obtained by treating the final cycle as a monotonic fracture test, evaluating K at the peak load.

Fatigue initiation length scales

Fatigue initiation length scales, ξ, were determined from cyclic crack growth resistance curves (ΔJ vs. Δa, or ΔK vs. Δa data) that exhibit two distinct power-law regimes. Each regime was fit to a linear model on a log-log scale, and ξ was defined as the crack extension at the intersection of these fits. Fitting ranges were determined algorithmically to ensure reproducibility. For the large crack regime, a linear fit was performed starting from the final data point, with the fitting window progressively extending towards earlier data points while R² ≥ 0.98; when R² fell below this threshold, the previous window size was retained. The minimum window size was the greater of 10 data points, or 5 % of the overall data. For the small crack regime, a linear fit was performed starting from the first data point, with the fitting window progressively extending towards later data; window sizes from 5 % to 40 % of the total data were tested, and the window maximizing R² was selected. The intersection was computed as ξ = exp[(C₂ - C₁)/(m₁ - m₂)], where (m₁, C₁) and (m₂, C₂) are the slopes and intercepts of the small crack and large crack fits, respectively.

Transmission electron microscopy

Transmission electron microscopy (TEM) imaging was conducted using a JEOL F200 transmission electron microscope operating at 200 kV. TEM micrographs were recorded using a OneView camera (Gatan Inc., USA). Cross-sectional TEM lamellae were prepared using an FEI Strata DB235 focused ion beam scanning electron microscope (FIB-SEM). Initial lamella preparation was performed using a 100 pA beam current at 30 kV, followed by low-voltage surface cleaning at 30 pA and 5 kV to minimize surface damage and amorphization.

Focused ion beam tomography

Focused ion beam tomography was performed using a Tescan S8252X FIB-SEM (TESCAN, Czech Republic) with a Xe⁺ ion plasma source at 30 keV, 300 pA with a 20 nm slice thickness. SEM images were acquired with a pixel size of 20 nm and a dwell time of 3 μs. Three-dimensional (3D) crack geometry was reconstructed from these image stacks and segmented on a slice-by-slice basis using the Segment Editor module in 3D Slicer. Initial segmentation was performed via intensity-based thresholding and manually corrected to exclude crack surfaces exposed in previous slices but not intersecting the current slice. To correct for beam drift, vertical translation for each sequential slice was estimated by cross-correlating one-dimensional profiles of the signed vertical intensity gradient. The calculated offsets were applied to the grayscale image stack using linear interpolation, and to the binary label map using nearest-neighbor interpolation with zero-padding to maintain the integrity of the segmented labels. Finally, a 3D median filter with a 3 × 3 × 3 voxel kernel was applied to the binary volume to smooth voxelated boundaries. Sequential color maps were generated by varying lightness in the CAM02-UCS perceptually uniform color space while preserving hue, with chroma scaled at lightness extremes to maximize contrast within the sRGB gamut.