Automated High‐Throughput Fatigue Testing of Freestanding Thin Films

Mechanical testing at small length scales has traditionally been resource‐intensive due to difficulties with meticulous sample preparation, exacting load alignments, and precision measurements. Microscale fatigue testing can be particularly challenging due to the time‐intensive, tedious repetition of single fatigue experiments. To mitigate these challenges, this work presents a new methodology for the high‐throughput fatigue testing of thin films at the microscale. This methodology features a microelectromechanical systems‐based Si carrier that can support the simultaneous and independent fatigue testing of an array of samples. To demonstrate this new technique, the microscale fatigue behavior of nanocrystalline Al is efficiently characterized via this Si carrier and automated fatigue testing with in situ scanning electron microscopy. This methodology reduces the total testing time by an order of magnitude, and the high‐throughput fatigue results highlight the stochastic nature of the microscale fatigue response. This manuscript also discusses how this initial capability can be adapted to accommodate more samples, different materials, new geometries, and other loading modes.


Introduction
Mechanical tests at conventional length scales are used to characterize the various mechanical attributes of materials, including elasticity, yielding, fracture, strain-rate sensitivity, temperature dependence, creep, and fatigue. However, challenges in sample preparation, load alignment, actuation, measurement, and material characterization can be problematic for the mechanical testing of small-scale materials. [1] Because of these difficulties, micromechanical and nanomechanical studies have often traded the DOI: 10.1002/smtd.202201591 interpretability of uniaxial tension for the experimental ease of simpler loading configurations, [2][3][4] such as nanoindentation, [5][6][7] bending, [2,8,9] and nanopillar compression. [10] To achieve uniaxial tension at the micro-and nanoscale, researchers have turned to costly but precise microfabrication approaches. [11][12][13] Materials of interest can be co-fabricated with their load frames to mitigate issues with sample preparation and load alignment. [14][15][16][17][18][19][20] This approach has facilitated fine control over tensile loadings, even within combined-loading environments. [21][22][23] However, most microfabrication-based approaches have employed temporallyserial testing of specimens, which is more time-consuming and less efficient at small-length scales [24,25] than at macroscopic ones. [26,27] At the microscale, fatigue testing poses an even larger challenge than monotonic tensile testing. The investigation of fatigue properties of smallscale materials has become extremely important due to the use of many flexible structures and thin films exposed to continuous and repeated loadings for a large quantity of cycles. For this purpose, microelectromechanical systems (MEMS) have been used under fatigue loadings in a variety of applications, ranging from high-temperature structures to mechanical/chemical sensors. [28][29][30][31] The main challenge with current small-scale fatigue characterization is that data is collected by testing a single specimen at a specified stress level or strain level. Moreover, the high-cycle fatigue response is stochastic in nature with a high degree of intrinsic sample-to-sample variation, necessitating multiple observations under each condition. Thus, fatigue tests must be conducted on numerous samples at a variety of stress levels or strain levels, rendering experimentation tedious for researchers, especially when a larger number of cycles is needed to reach fatigue failure. Additionally, in situ testing, such as inside a scanning electron microscope (SEM), [32][33][34] synchrotron X-ray system, [35,36] or transmission electron microscope (TEM), [37] is often prohibitively expensive due to the requisite instrument time. These challenges highlight the value of new techniques that can increase the throughput of microscale fatigue experimentation, improving its efficiency and enabling faster exploration.
At the macroscale, Zhou et al. [38] developed a high-throughput experimental procedure to capture rare-event failure of materials. They developed an electric displacement stage that simultaneously strained up to 1000 additively manufactured samples. Samples were imaged with a traditional camera as they were cycled until failure. This approach enabled the identification of rare-event failures with high statistical accuracy. There are notable prior efforts to extend high-throughput testing to the microscale. Oellers et al. [14] developed a photolithographic, high-throughput process to fabricate thin film tensile-test structures, which were tested sequentially inside an SEM to obtain mechanical properties. Additionally, Burger et al. [8] studied high-throughput, high-cycle fatigue damage in metallic thin films coupled to vibrating Si cantilevers. The surface stress at the thin film would vary along the position of the cantilever beam, so they were able to quantify fatigue damage along a wide range of stress amplitudes in a single sample. However, none of these approaches was able to both avoid the tedium of sequential testing and the potential substrate coupling.
To advance the state-of-the-art, the current work presents a new methodology for the high-throughput fatigue testing of freestanding thin films. This methodology features a microfabricated, MEMS-based Si carrier with an array of samples that can be simultaneously and independently tested under mechanical loadings. The gauge sections of the samples are freestanding to avoid confounding effects from an attached substrate. Automated in situ SEM fatigue testing of this array can efficiently quantify statistical variations in the degradation mechanisms at the micro-and nanoscale. This approach significantly reduces the preparation time, tool time, and funding when compared to traditional, sequential fatigue tests. It also allows for the investigation of the mechanistic nature of stochasticity in fatigue, which is more prominent than in other mechanical environments. To demonstrate this methodology, the fatigue behavior of freestanding nanocrystalline Al thin films was characterized.

Design
The main vision of the design was to develop a MEMS-based structure that would enable high-throughput micromechanical tests with samples loaded simultaneously. The idea emerged from the concept of metamaterials. [39][40][41] Instead of using a periodic geometry to withstand a mechanical load, the design envisioned in this work employs a periodic geometry to apply a mechanical load to multiple specimens simultaneously. Figure 1 depicts a solid model of the proof-of-concept design. The geometry was designed via Cubit 15.5.0 by Sandia National Laboratories [42] and evaluated via Abaqus/CAE 2017 by Dassault Systèmes. [43] Each unit cell, which is similar to isolated tensile testers in prior works, [17,44,45] strained a dogbone-shaped, freestanding thin film. By tessellating this unit cell into a grid, the resulting device can simultaneously and independently test 12 freestanding thin films, although larger arrays could be developed in the future.
This device features three areas: 1) the array of freestanding films and alignment springs in the middle, 2) grip regions at the left and right, and 3) temporary supports at the top and bottom. These temporary supports protect the device during handling and are removed just prior to mechanical testing. Loads can be applied either via pin-loading at the holes or through clamploading at the grip surfaces. The widths of these grips were chosen to ensure that the applied load would distribute evenly across the sample array and therefore promote equal loading of each thin film. Further, the geometry of the sample array itself, especially the springs, was chosen to ensure that the features in the individual unit cells would deform elastically, even in the case that the thin films deformed plastically. Otherwise, film defects stochastically induced during processing could complicate the evaluation of the loading conditions and/or impede the mechanical characterization of the films. Finally, the single crystal Si carrier was chosen to be three orders of magnitude thicker than the metal layer (i.e., ≈300 μm vs ≈320 nm) to promote independent deformation of the freestanding films within the metal layer. A corollary to Saint Venant's principle [46] suggests that the compliant metal film should be at least ten times thinner, and preferably 100 times thinner than the Si carrier substrate; in the current study, the ratio of film to substrate thickness was even smaller at 1:1000, to provide conservative assurance that the film would not influence the carrier. Hence, even if some unit cells deformed undesirably during processing or testing, the remaining unit cells would still be functional for subsequent testing.
To ensure that the mechanically applied forces and deformations at each specimen location were appropriately sized and independent and to ensure that the deformation of the Si frame was negligibly affected by the presence/deformation of the dogbones, two sets of displacement-controlled, elastic, finite-element analysis (FEA) were performed on the device. The first set was performed on the entire design to evaluate the local deformation of the dogbones. The second set was performed on only the Si carrier to ensure that it would deform uniformly and to evaluate the stress concentration on the springs. As would be the case for an experimental test, the temporary supports (shown in Figure 1) were removed for the FEA. In Cubit, both solid models were meshed with 10-node tetrahedra, which captured the bending-dominated deformations in the springs of the carrier. The dogbones in the first set of FEA exhibited a finer mesh to capture their deformation in detail. For boundary conditions, the outer surface of one pin-loading hole was fixed and the displacement of the other hole was set to 20 μm. The simulations were run in Abaqus with Al elastic properties of E = 70 GPa and = 0.35 [47] and Si elastic properties of E = 170 GPa and = 0.27 [48] (corresponding to the [110] direction of the (001) Si wafer used in fabrication (see Microfabrication section)). While Si is slightly anisotropic, the isotropic modulus is a reasonable approximation in this case. [49] Figure 2b,c shows the resultant maximum principal strain and the stress state ( on the dogbones under these conditions. Since the dogbones are comparatively thin (≈320 nm), the out-of-plane stress (S 33 ) is zero. Regarding the simulation without dogbones, Figure 2d shows that this conditions yielded a maximum von Mises stress of 150 MPa, which was comfortably below the expected values for the strength of microfabricated silicon. [50] This quasistatic simulation was sufficient for ensuring that the Si carrier would have nearly infinite fatigue life at these stress levels. [51] These results demonstrated that the deformation of the Si frame was identical at each of the 12 dogbone locations, even when dogbones were stressed far beyond their elastic limit. Modeling the dogbones in this fashion functioned as a conservative test of the Si deformation. While the FEA ensures the integrity of the design, actual mechanical tests will feature even smaller strains/stresses.

Microfabrication
This section describes the steps required for the microfabrication of the high-throughput device, which is depicted in Figure 1. The detailed steps, summarized in Figure 3, can be separated into three subsets: steps 1-4 for producing a metal-coated wafer, steps 5-8 for patterning the metal coating on the top of the wafer, and steps 9-13 for etching the Si from the bottom of the wafer to yield freestanding films.

Producing the Metal-Coated Wafer
Step 1 started with a double-side-polished (DSP), 2″ p-type (001) Si wafer with a thickness of 279 ± 25 μm. This type of wafer was necessary for the upcoming through-wafer etching. To establish a reference point for later stress measurement, step 2 measured the curvature of this wafer.
Step 3 deposited ≈320 nm of Al via DC magnetron sputtering (KJL PVD-75) over a total of 6500 s.
Step 4 quantified the deposition-induced average film stress (i.e., −175 MPa ± 25 MPa) via Stoney's equation [52] and a second curvature measurement. This slight compressive stress was ideal because it would avoid all preloading preceding the mechanical testing.

Patterning the Metal Coating on Top of the Wafer
To pattern the Al layer, step 5 first patterned photoresist atop this Al using a Heidelberg Maskless Laser Aligner (MLA) 150. This process included: i) spin-coating AZ9260 photoresist onto the Si wafer at 3000 rpm ii) removing the solvent in the photoresist by baking at 110°C for 90 s, iii) writing the desired pattern via the MLA with a laser exposure energy of 700 mJ cm −2 , and iv) removing the exposed photoresist by soaking in Microposit Developer Concentrate until visibly clear (≈60 s). This pattern contained 15 devices and featured four alignment markers (for facilitating the upcoming backside patterning) on the 2″ wafer. The tensile axes of all 15 devices were aligned with the major flat of the (001) wafer, following the [110] direction. Step 6 removed the Al not covered by the photoresist via soaking in a liquid Al etchant type A until visibly clear (≈8.5 min). Because the etchant highly selected the Al over the Si, the thickness of the Al layer could be measured via profilometry after stripping the photoresist in steps 7-8. This thickness value was required to validate the average film stress calculated in step 4 above.

Etching the Si from the Bottom of the Wafer to Yield Freestanding Films
In preparation for a through-wafer etch from the backside, step 9 patterned the backside Si with a photoresist via the MLA with backside alignment. This pattern was identical to the one in step 5 except for the absence of the dogbones, alignment markers, and sample numbers. To withstand prolonged exposure to BOSCH etching (which removes more photoresist than the wet etching of step 6), step 9 used a thicker photoresist than step 5. Specifically, step 9 spin-coated AZ9260 at 1300 rpm (rather than at 3000 rpm) to deposit ≈10 μm of photoresist. The bake at 110°C for 90 s and the laser exposure energy of 700 mJ cm −2 were retained but development required an additional 30 s. Because a BOSCH etch would shut down if the etch were allowed to go through all inserted materials, step 10 mounted the Al-Si wafer onto the center of a 4″ sapphire wafer. This process comprised: i) thinly spincoating the sapphire wafer with AZ9260 at 5000 rpm, ii) immediately pushing the silicon wafer (backside up) onto the photoresist, and iii) baking at 110°C for complete removal of the solvent in the photoresist. The sapphire wafer was preferable to a 4″ Si wafer because it reduced the chemical loading on the BOSCH etch.
Finally, step 11 applied a through-wafer BOSCH etch, also known as deep reactive ion etch (DRIE), via a Plasmatherm model 770SLR to yield freestanding Al films and separate the devices from the Si wafer. The BOSCH etch in this outdated model was weakest in the center of the wafer. Hence, the outermost samples were iteratively examined to detect the point of breakthrough, which typically occurred soon after 700 cycles of 15-s BOSCH etching. Additionally, the etch had a heterogeneous angular distribution, so the wafer was manually rotated 90°every 200 cycles. After 600 cycles, the final orientation of the wafer was retained and the etch depth was checked more frequently to avoid over-etching, which often destroyed devices. Issues with yield were almost exclusively due to difficulty with etching through the Si wafer. Most importantly, etch heterogeneity resulted in a timeconsuming verification process and lower yield. Modern state-ofthe-art etch tools would mitigate these issues and improve yield, hence enabling devices with larger numbers of samples. Typically, the BOSCH etch required a total of 750 to 800 cycles. At this point, the devices had separated from the Si wafer but were still attached to the sapphire wafer via photoresist. To remove this final attachment, step 12 soaked the entire wafer stack in the proprietary N-methly-2-pyrrolidone-based solvent "Remover PG" at room temperature overnight. To remove residual photoresist, step 13 briefly soaked each device in acetone, methanol, and deionized water and then performed an oxygen plasma clean at 150°C for 120 s.
In this proof-of-concept study, with planning and ample tool availability, the time needed for the microfabrication of 15 devices is around 1 day. This microfabrication procedure is easily scalable with the prospect of producing a large number of devices without a substantial cost increase, as is the case for other microfabricated small-scale testing structures. [53,54] For example, if a 12″ wafer instead of a 2″ wafer was used, hundreds of these devices could be fabricated for similar costs. Figure 4 shows SEM images (acquired on a JEOL IT500HRLV) of a representative microfabricated device. Figure 4a shows the array of springs, grips, and dogbones of the Si frame with the thin Al layer deposited on top, and Figure 4b shows a unit cell. Figure 4c shows a single Al dogbone and Figure 4d shows a high-magnification image of the freestanding Al material. The dogbones exhibited a slight sag after the etching of the Si layer,  demonstrating that the deposition-induced compressive stresses relaxed at this region. A single device usually yielded 6-12 intact dogbones for subsequent mechanical testing.

High-Throughput Device and Material Characterization
To characterize the grain structure of the thin Al freestanding film, transmission kikuchi diffraction (TKD) was performed via an Oxford Instruments Symmetry electron backscatter diffraction (EBSD) camera in a Zeiss Supra 55VP SEM. Figure 5a,b shows the band contrast and the inverse pole figure (IPF) map of the plane-view microstructure, respectively. The Al dogbone exhibited randomly oriented grains with an in-plane mean average grain diameter of ≈120 nm. Additionally, atomic force microscopy (AFM) via an Oxford Instruments Asylum Research MFP-3D characterized the surface roughness of the Al layer, highlighting hillock grains that protruded from the surface and were ≈100 nm taller than the average grain height.

Monotonic Tensile and Fatigue Testing with Automated SEM Imaging
In preparation for the mechanical testing of these devices, the temporary supports (shown in Figure 1) were removed via a femtosecond-pulsed Ti:Sapphire laser to prevent damage to the Al dogbones. Removing the supports with mechanical forces was avoided because the resulting forces propagating through the Si carrier were observed to damage the Al dogbones, rendering the devices unusable for subsequent mechanical testing. After the temporary supports were removed, each device was carefully lifted out and placed into a mechanical testing stage, minimizing lateral forces that could pre-strain the dogbones. The mechan-ical testing stage was a custom-built piezoelectrically actuated, miniature load frame [34,55] used for tensile and fatigue testing. All mechanical tests were conducted in situ, under a vacuum environment (<10 −6 Torr) in a JEOL IT500HRLV SEM operated at 10 kV.
NextAn in situ calibration test was performed to relate local strain to far-field displacement. The device was subjected to uniaxial tension at discrete displacement intervals of 0.5 μm (corresponding to a 0.125 μm displacement in each Al dogbone), and a 20-s secondary-electron image was acquired of the gauge section of a dogbone sample at each interval. The captured images were analyzed via digital image correlation (DIC) software (VIC2D) to calculate the strain a single dogbone experienced at each actuated displacement interval. From this analysis, the displacement of the entire device was related to the local strain in each dogbone (see the plot of DIC monotonic tensile tests in Supporting Information). The stiffness of the device was measured to be 3 mN μm −1 via these monotonic tests. While the entire highthroughput device was actuated, the Si carrier only experienced elastic behavior, even when a dogbone ruptured.
Finally, in situ SEM fatigue tests were performed under displacement control at a frequency of 1 Hz with a mean displacement of 1.4 μm. The controlled displacements of the Al dogbones associated with the tests had maximum values ranging from 1.7 to 2.3 μm, corresponding to maximum strains ranging from 1.7% to 2.1%. Depending on the overall duration, the tests were interrupted every 500 or 1000 cycles for SEM imaging, while holding the dogbones at the mean displacement. During each interruption, a custom Python script (shown in the Supporting Information) would control the SEM to sequentially move the SEM stage www.advancedsciencenews.com www.small-methods.com to the location of each dogbone and acquire multiple plane-view images at different magnifications of the gauge section of each dogbone. The automatic focusing, stigmation, and brightnesscontrast features of the SEM, also activated via the script, permitted the images to be collected consistently. The fatigue tests were fully automated (with the exception of the occasional SEM beam shift to compensate for sample drift and re-center images at the gauge section of the dogbones) and ran until all dogbones failed or until the script was stopped. Fatigue test durations ranged from a few hours up to 3 weeks of continuous testing. When dogbones failed, the mean displacement of other dogbones remained constant. With the acquired SEM images, the fatigue life of each dogbone sample was determined, and corresponding imagery of the crack initiation and crack propagation processes was compiled.

Figure 6a
shows the strain-life curve of the fatigue behavior of the Al dogbones. The fatigue life (N f ) was characterized as a function of applied strain amplitude ( a = max − mean ). Strain amplitudes ranged from 0.25% to 0.62%, corresponding to fatigue lives (N f ) ranging from <500 to >10 6 cycles. The applied mean strain ( m ) was 1.5% as calculated via the DIC calibration mentioned above. The right-pointing arrows represent run-out specimens (which did not reach fatigue failure), and the numbers in parentheses represent the number of samples that failed at the specified number of cycles. In Figure 6a, the red solid line represents the mean average fatigue behavior of nanocrystalline Al in the strain amplitude range studied, and the shaded region represents the 95% prediction interval (PI) of the fatigue lives. The average fatigue behavior was obtained via linear regression on the logarithmic strain-life curve. The 95% PI estimates the range in which a future sample would reach fatigue failure with a 95% probability. With only four automated, high-throughput, in situ SEM tests, the fatigue behavior of nanocrystalline Al was efficiently characterized while also assessing the stochastic variability of fatigue lives and capturing the morphological evolution of all samples during the tests. Figure 6b compares the fatigue behavior of nanocrystalline Al (collected using the current approach) with the fully reversed, traditional fatigue behavior of polycrystalline Al. The conventional fatigue behavior of structural Al is shown according to the Coffin-Manson Equation (Basquin coefficient b = −0.102 and Coffin-Manson coefficient c = −0.645). [47] Both behaviors follow a similar trend, but the nanocrystalline Al curve was acquired at a mean strain of 1.5%. The mean stress was estimated to be ≈130 MPa via the previously-known uniaxial stress-strain behavior of Al thin films. [44,56] Using the Morrow-mean-stresscorrection equation [47] on fatigue lives, the conventional fatigue behavior was shifted. The corrected fatigue behavior falls within the 95% PI. At high total strain amplitudes, the mechanical behavior of nanocrystalline Al deviates from the polycrystalline behavior, hence highlighting a reduced low-cycle fatigue resistance that is often observed in nanocrystalline metals due to their limited resistance to crack propagation. [57] Figure 7a-d shows four before-and-after comparisons between the untested and failed dogbones of an in situ fatigue test performed on a single device, in which 12 dogbones were tested at a strain amplitude of 0.44%. The fatigue lives ranged from 1000 cycles up to a run-out specimen that did not fail after 3 × 10 5 cycles, hence highlighting the stochastic variability of fatigue lives under the same applied conditions. This large variability in fatigue life is expected in nanocrystalline metals [58][59][60] due to the intrinsic stochasticity in microscale fatigue owing to the heterogeneous nature of the weakest link features that govern failure. [61,62] However, such variability is rarely quantified to such an extent due to the time-consuming nature of traditional microscale fatigue experiments. A video of a selected in situ test is included in the Supporting Information. The video is time-lapsed, where each frame represents a progression of 1000 cycles.
The automated in situ SEM tests allowed for the evaluation of morphological changes along the Al dogbone's surface during testing. For example, Figure 8a-d shows the crack initiation and crack growth on a dogbone tested at 0.30% strain amplitude. These images can be used to quantify crack initiation life and crack propagation life and can also reveal microstructural evolutions, such as the microcracks shown in the figure. In this particular dogbone, the initiation life was 2.7 × 10 4 cycles (corresponding to 84% of its fatigue life), and the propagation life was 5 × 10 3 cycles (16% of its fatigue life). The crack propagation rate as a function of total strain amplitude could be determined from these experiments. With the current imaging rate, only a mean crack propagation rate of da/dN = 4 × 10 −9 m/cycle could be estimated because only two images were obtained at intermediate crack lengths. In future automated experiments, edge-tracking algorithms could be used to identify when cracks nucleate, and imaging could correspondingly be adjusted to detail the crack propagation. The mean crack propagation rate in this sample is commensurate with propagation rates found in other submicron thin films. [63][64][65] Additionally, in the images preceding crack nucleation, there were no observable salient features at the top surface of the dogbones (e.g., slip bands or intrusions/extrusions), consistent with previous observations in the fatigue damage of nanocrystalline nanometer-thick films. [66] However, previous studies have observed clear slip bands and extrusions in micrometer-thick films with nanocrystalline and ultrafine grains. [9] In the future, thicker dogbones could be deposited via the current methodology to investigate extrusion formation with high-throughput experimentation.
The in situ fatigue tests also showed the crack propagation path on the dogbones. The path of the crack in all failed samples exhibited significant tortuosity, reminiscent of an intergranular fracture surface in bulk metals. Figure 9a shows a dogbone subjected to a = 0.44% with N f = 3.9 × 10 4 cycles at its untested state and just after failure. The trace of the crack path (yellow solid line) was added to the image of the untested dogbone to highlight surface features that may have influenced the crack path. As highlighted by the yellow arrows, the crack tended to follow grains that protruded from the surface of the film. These protruding surface grains (hillocks) are commonly observed in deposited Al thin films. [67,68] The hillocks served as stress concentration sites that facilitated crack propagation. Figure 9b shows an oblique view of the fracture surface of a dogbone tested at a = 0.44% with N f = 1 × 10 3 cycles. This postmortem image also highlights the significant crack tortuosity and hillocks along the path of the crack. Hence, these analyses confirm that the stochastic failure within the gauge section of the Al dogbones is associated with these hillocks formed during deposition, rather than any other manufacturing-induced defect. The topography of the Al dogbones presented here is typical of the actual surface topography of many deposited films and has significantly influenced the fatigue crack behavior.
An additional in situ stress relaxation test was performed to verify the influence of time-dependent behavior on the failure of the dogbones. The dogbones were held at the maximum strain studied (2.1%) for 8 h, and dogbones did not fail. Fatigue results of dogbones cycled at a strain amplitude of 0.6% and mean strain of 1.5% showed fatigue failure after 10 000 cycles (≈3 h). These results indicate that fatigue failure in the dogbones was not substantially influenced by time-dependent stress relaxation.

Discussion
This study presented an efficient, adaptable methodology for the high-throughput fatigue testing of thin films. Experiments performed on freestanding, nanocrystalline Al thin films revealed the stochasticity of fatigue lives and the nature of the crack evolution at several strain amplitudes.
Compared to traditional micromechanical testing, this approach offered three key benefits. First, the microfabricated devices significantly reduced the sample-alignment time. The alignment of samples for microscale testing normally requires meticulous handling and preparation for each individual test. This approach allows all samples to have proper and identical alignments by integrating all samples within their own Si carrier. Second, the automated high-throughput fatigue testing dramatically reduced the amount of time necessary to collect the same amount of data under a myriad of traditional fatigue tests. While current testing was limited to a 1 Hz loading rate by the current actuator, future systems could employ faster actuators to further accelerate testing. Third, these tests also maximized the efficient use of instrument (SEM) time. For example, in the longest in situ SEM test, the SEM ran continuously for 3 weeks, collecting detailed fatigue data for a number of samples. In this first demonstration, the amount of time saved in testing was an order of magnitude, given that there were up to 12 samples tested independently and in parallel per device. Building on this initial concept, future devices could be developed to accommodate hundreds of fatigue samples, resulting in a commensurate reduction in data acquisi-tion time. Moreover, the geometry of the Si carrier could be modified to exert different strain levels on different samples in a single device. Likewise different sample shapes could be designed to exert different loading conditions (e.g., varied notch geometries on different dogbones in a single device).
Compared to the state-of-the-art, high-throughput, micromechanical testing techniques, [8,14] the approach presented in this work also offers a number of advantages. Fatigue samples here were tested independently and simultaneously, gathering statistically significant data faster than many high-throughput techniques that use serial testing. The approach presented here focused on the mechanical behavior of thin films from a purely mechanical point of view, and therefore, the freestanding samples decrease the level of complexity in experimental analysis and eliminate concerns about the influence of a substrate in fatigue behavior. However, for engineering applications requiring a coupled substrate and thin film, other techniques may be more suitable. Additionally, this methodology avoids the use of a focused ion beam (FIB) for sample preparation, eliminating any concerns of FIB-induced damage that could alter the structure, and therefore, the mechanical behavior of small-scale materials.
However, this methodology also exhibits some limitations. In traditional single-sample fatigue testing, fatigue lives are commonly determined by a significant decrease in load due to fracture. Since the thin films on the Si frame are fabricated to be tested in parallel, it is not possible to directly measure the load on individual samples in the current configuration, although the strain on each sample was determined independently via the global displacement measured at the grips, local DIC of the dogbones, and a displacement-strain calibration experiment. Consequently, the stress amplitude and mean stress in a unique sample could not be determined. However, in future iterations of this work, flexible struts that act as load sensors can be added to the grips in order to estimate loads by measuring the deflection of the struts. [69,70] With the current approach, the fatigue lives of each sample are instead found through SEM images, which are only acquired at determined intervals. Therefore, it is necessary to combine this high-throughput testing methodology with a highmagnification imaging capability. Additionally, because of the dimensions of these devices, this approach would only be suitable for investigating submicron-and micron-thick films.
The current work serves as a proof-of-concept of a new methodology for high-throughput fatigue testing of thin films and a first demonstration with 320 nm-thick nanocrystalline Al dogbones as an example of study. This demonstration is not a comprehensive evaluation of the new capability, nor is it in its final stage of development. The current methodology could be modified to deposit various thicknesses of Al films (up to a few microns thick films) to study the influence of film thickness on extrusion formation and crack propagation under fatigue loadings. With additional design modifications, the length, width, and shape (e.g., notches) of the freestanding films can also be changed. These films can also be annealed over a range of temperatures to modify the grain sizes and residual stresses and the microstructure effect can be easily studied. In future iterations of the methodology, the substrate design can be modified to accommodate hundreds of samples per Si device, different sample geometries, and different loading types (e.g., inplane shear, out-of-plane bending, and combined tensile-shear loadings). The material composition of the film can also be varied according to the deposition capability and substrate adhesion. While the focus of the present study was on fatigue behavior, the methodology also conceptually allows the investigation of thin film behavior under different stimuli, such as radiation, heating, or environmental. The technique is also conceptually amenable to different material characterization techniques, such as TEM, AFM, and X-ray scattering. Additionally, the methodology allows for high-throughput investigation of combined stimuli, such as radiation-fatigue, [71] creep-fatigue, thermo-mechanical fatigue, and environmental fatigue. [51] Overall, this new high-throughput approach has the potential to advance the state-of-the-art for statistically assessing the effects of a library of parameters (microstructure, material, geometries, loadings, etc.) on small-scale material behavior. Figure 9. a) Crack propagation path in an Al dogbone tested at a = 0.44% with N f = 3.9 × 10 4 cycles. The yellow arrows highlight hillocks on the surface of the dogbone, which facilitate crack propagation. b) Fracture surface of another dogbone tested at a = 0.44% with N f = 1 × 10 3 cycles demonstrating the crack tortuosity and the hillocks along the crack path.

Conclusion
This work developed a new methodology for the high-throughput fatigue testing of freestanding materials at the microscale. This methodology featured the microfabrication of a high-throughput Si carrier that supports the simultaneous and independent fatigue testing of an array of samples. The combination of this methodology with the automated in situ SEM testing enabled efficient characterization of the fatigue response of thin films at different applied strain amplitudes with tests that would run automatically for up to 3 weeks at a time. As a proof-of-concept, this approach efficiently assessed the fatigue behavior of nanocrystalline Al thin films. Results highlighted stochasticity in fatigue lives, where lives differed by up to 2.5 orders of magnitude for samples tested under nominally identical conditions. Results also indicated that fatigue lives were heavily dominated by their initiation lives, yet no fatigue-induced topographic extrusions were found at the crack initiation sites. Fatigue cracks exhibited significant tortuosity and tended to follow hillocks on the surface of the dogbones. The current methodology offers many benefits when compared to traditional, sequential fatigue testing of small-scale materials and other high-throughput techniques. The methodology can be used to investigate a myriad of parameters in material behavior and can be further expanded to accommodate more samples, different materials, varied geometries, and other loading types and rates.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.