Diffraction Line Broadening as a Second Witness for the Indentation Size Effect

Push a sharp diamond tip a few tens of nanometres into a metal and the metal looks harder than it does at larger depths. Two explanations fit this "indentation size effect" equally well. One says the extra resistance comes from a thin, rough surface skin: the indenter crushes tiny surface bumps, and that plastic work acts like an effective surface tension, the apparent surface stress $f$. The other says the extra resistance comes from the bulk: the tip's shape forces the crystal to store extra dislocations, the line defects that carry plastic flow, below the contact. Both explanations predict hardness rising roughly as $1/h$ with depth $h$, so the load-versus-depth curves that indentation produces cannot tell them apart. The two pictures do disagree about where the deformation sits. The surface picture puts it in a skin about as thick as the roughness. The bulk picture spreads it over a volume that scales with the contact size.

This note proposes measuring that location directly through strain-induced inhomogeneous broadening: when different parts of a crystal sit at different elastic strains, a spectral or diffraction peak smears out, because each region contributes a slightly shifted line. Spectroscopists usually treat this broadening as a nuisance. Here it becomes the signal. We indent dense arrays into surfaces prepared to separate roughness from subsurface damage, record X-ray diffraction peak widths at several probing depths, and ask which mechanics model, fitted only to the hardness curves, correctly predicts the broadening profile it never saw. The model that does is the physically correct account of shallow-depth hardness. The same forward model can then predict interface-induced broadening in bonded thin films, where it currently limits spectroscopic resolution.

Core Claim

The hardness curve $H(h)$ alone does not identify the mechanism. The joint data of $H(h)$ plus depth-resolved diffraction broadening under the same indents does. Concretely, a surface-dissipation model and a bulk strain-gradient model can be tuned to the same $H(h)$, but they then make different, parameter-free predictions for how peak broadening varies with X-ray penetration depth. The surface model predicts broadening concentrated within roughly the roughness amplitude and insensitive to deeper probing. The bulk model predicts broadening extending to depths of order the contact radius and scaling with indentation depth. A second claim follows from the first. If apparent surface stress is real boundary dissipation, then $f$ measured on one surface preparation should be predictable from measured topography alone, with bulk parameters frozen.

Prior Work and Its Role

The surface side is defined by Xu and Zhang's Surface Effects on Nanoindentation. That paper splits indentation work into bulk and surface parts and fits $H = H_0 + g f / h$, where $H_0$ is the large-depth hardness and $g$ a tip-geometry factor, across metals, ceramics and polymers. Its striking result is that changing only the polish on (111) Cu moves $f$ from 18.9 to 60.0 J/m². Zhang, Xu and Zhao's The role of plastic deformation of rough surfaces in the size-dependent hardness supplies the mechanism: a bearing-ratio model of asperity flattening, where the bearing ratio is the fraction of nominal area actually in contact at a given height. They tested it on Sn-Pb and Teflon polished with graded abrasives, and they already used Raman and XRD for surface residual stress. Both papers motivate this proposal. Both also share its weakness: $f$ is inferred inversely from hardness curves and never located in space. Mechanical polishing also changes subsurface damage along with roughness, so the polish result is confounded.

The bulk side is represented by Indentation of a hard film on a soft substrate: Strain gradient hardening effects. It explains the size effect in a 640 nm W film entirely through conventional mechanism-based strain-gradient plasticity, a continuum theory in which plastic-strain gradients raise the flow stress through geometrically required dislocations. It assumes a smooth contact and calibrates an internal length $l_W = 0.23$ μm. Energetic Approach to Large Strain Gradient Crystal Plasticity gives a variational large-strain version with no contact-dissipation term. These papers supply our competing hypothesis and its implementation. Our proposal does not reject them; it adds an observable that can falsify them.

Numerical Simulation of the Residual Stress at the Interface between Thermal Barrier Coating and Nickel-Based Single-Crystal Superalloy Based on Crystal Plasticity Theory provides a method we reuse: crystal-plasticity finite elements, in which plastic flow is resolved on individual slip systems, applied to explicitly rough interfaces. It reports that crystal plasticity raises interface tensile stress up to 2.5 times over creep-only models. It also shows that a single-harmonic sinusoid already produces strong peak/trough concentrations. We replace its idealized sinusoid with measured three-dimensional topography and add a boundary dissipation term, which it lacks.

On the broadening side, Quadrupolar effects on nuclear spins of neutral arsenic donors in silicon shows that mismatch strain in silicon bonded to sapphire produces quadrupole splittings of about 10–31 kHz, which inhomogeneous strain then smears. Strain-Induced Enhancement of the Charge Density Wave in the Kagome Metal ScV6Sn6 shows a bending-tuned collective mode moving from about 1.34 to 1.39 THz, likewise blurred by non-uniform strain. Both motivate the claim that broadening encodes strain distributions worth modelling rather than discarding. Neither treats interface contact mechanics. That is the extension we target last, after the method is validated where the ground truth is cleaner.

Research Program

The program has three parts: a specimen design that breaks the polish confound, a measurement of broadening that resolves depth, and a forward model that turns each mechanics hypothesis into a predicted peak width.

Specimens. The first material is (111) single-crystal Cu, chosen because the file's only hard numbers for $f$ live there. We build a factorial design that separates topography from subsurface damage. A mechanical polish gives a rough, damaged surface. The same polish followed by a low-temperature recovery anneal is intended to give a rough surface with reduced subsurface dislocation content. Before and after the anneal, we check with optical profilometry (roughness and bearing ratio) that the topography is unchanged. An electropolish gives a smooth, undamaged surface. Electropolish followed by low-energy ion irradiation adds near-surface defects with minimal roughening. That last cell is the least certain to work cleanly, and we report it as such. The second material is a 640 nm W film on sapphire, matching the specimen type for which the strain-gradient baseline is already calibrated. A bonded film is also the geometry relevant to the spectroscopy papers.

Broadening measurement. On each surface we place arrays of Berkovich indents at fixed target depths spanning roughly 30–300 nm, which straddles the reported critical depths of 75 and 111 nm. Spacing is set at several contact radii so that indent fields do not overlap. Broadening is an ensemble quantity, so the array supplies enough diffracting volume. We use grazing-incidence XRD at a ladder of incidence angles, which sets X-ray penetration from tens of nanometres to microns, to record peak width as a function of probing depth for each indentation depth. Indent-free regions of the same surface are measured as the reference. High-resolution EBSD (electron backscatter diffraction, which maps local lattice misorientation) on focused-ion-beam cross-sections gives an independent dislocation-density map for a subset of indents.

Forward model. Each hypothesis is run as a crystal-plasticity simulation of indentation. The surface model uses measured topography with a boundary dissipation term set by the bearing-ratio construction and no gradient hardening. The bulk model uses a smooth surface with strain-gradient hardening. The joint model includes both. From each simulated elastic strain field we compute the predicted peak width. We project the strain onto the diffraction vector and weight each element by $\exp(-z/\Lambda)$, where $z$ is depth and $\Lambda$ the penetration depth at a given incidence angle. The spread of the resulting lattice-spacing distribution is the predicted broadening $\beta(\Lambda, h)$. Instrumental and indent-free widths are removed by convolution, not subtraction.

Calibration and surrogate. Every model is calibrated on $H(h)$ only. $\beta$ is held out. Inverting over many measured topographies is expensive, so we may train a neural-operator surrogate in the style of Learning the Stress-Strain Fields in Digital Composites using Fourier Neural Operator, which maps 2D microstructures to full stress-strain fields. That work was trained only on synthetic binary composites under tension, so any surrogate is validated against direct simulations on held-out topographies before use.

Baselines

The naive reference is the classical continuum answer: a depth-independent hardness $H_0$ with no size effect. It predicts broadening equal to the broadening a smooth-surface, gradient-free plasticity simulation produces. Every hypothesis must beat it on $H(h)$; it is included mainly to scale the broadening excess that needs explaining.

The strongest existing surface-side method is the $H$ versus $1/h$ fit of Surface Effects on Nanoindentation. It reports $f = 60.0$ J/m², $H_0 = 0.662$ GPa and $h_c = 75$ nm for mechanically polished (111) Cu, and $f = 18.9$ J/m², $H_0 = 0.718$ GPa and $h_c = 111$ nm for electropolished Cu. It also reports a higher correlation coefficient than dislocation-only fits on (100) Al, though the file gives no numerical $R^2$. We refit it on our own $H(h)$ data with the same depth window and the same exclusion rules, and we first check that we reproduce those values within scatter on matched preparations.

The strongest bulk-side method is the strain-gradient finite-element model of Indentation of a hard film on a soft substrate: Strain gradient hardening effects. For the W film it starts from that paper's calibrated $l_W = 0.23$ μm and 0.88 GPa lattice friction stress. For Cu, its internal length is fitted per preparation, with the same tuning budget as the surface model: an equal number of free parameters and the same optimizer and number of restarts.

For peak broadening under indent arrays, no reported result was found on Lacuna, so the broadening comparison has no published number to beat. The criterion is held-out predictive error. The two ablations that isolate the key ingredient are the joint model with the boundary dissipation term removed and the joint model with gradient hardening removed. Beating the baselines means the following: with comparable $H(h)$ residuals, the winning model's predicted $\beta(\Lambda, h)$ falls within measurement uncertainty across all incidence angles, while the competing model misses systematically in the direction its geometry dictates.

Why not just…?

Why not just compare which formula fits the hardness curve better? That is what the existing literature did. The (100) Al correlation-coefficient advantage reported in Surface Effects on Nanoindentation is real but cannot be decisive. Both families have one slope-like free parameter, and both produce near-$1/h$ behaviour at shallow depth. A goodness-of-fit difference on one curve mostly reflects fitting-window choices. Our design reports these fits too, but treats them as the starting point, not the verdict.

Why not just image dislocations under the indents with EBSD or TEM? This is the most serious alternative, and we include it rather than argue it away. Its limits are specific. The relevant region sits below about 100 nm. Ion-beam cross-sectioning itself damages that layer. Misorientation maps also see stored dislocations, not the elastic strain spread or the dissipated work. EBSD alone therefore cannot easily distinguish "asperity crushing that left little stored damage" from "a modest gradient field." We run HR-EBSD on a subset and test it as a standalone discriminator. If it separates the factorial cells as cleanly as broadening does, we will say so, and the contribution narrows to the forward model and the topography-predicts-$f$ test.

Why not just vary polishing grit and watch $f$ change? The role of plastic deformation of rough surfaces in the size-dependent hardness did exactly this with graded abrasives. Coarser grit raises roughness and subsurface damage together, so a bulk model can absorb the change by raising initial dislocation density. The recovery-annealed and ion-damaged cells exist to break that coupling, and the decisive test below depends on them.

Decisive Checks

The first check is out-of-sample prediction of broadening. Fit each model to $H(h)$ on all cells, then score its predicted $\beta(\Lambda, h)$ against measurement. The surface hypothesis is falsified if excess broadening grows with penetration depth in proportion to contact radius. The bulk hypothesis is falsified if the excess is confined to the shallowest probing angles and tracks measured roughness rather than indentation depth.

The second check is transfer across preparations. Freeze bulk parameters on the electropolished cell, feed in only the measured topography of the rough cells, and predict their $f$. The prediction should approach the reported jump toward 60 J/m² in the rough-damaged cell. It should also persist in the rough-annealed cell if the jump is due to topography, and vanish there if it was due to damage. The joint model passing both checks where single-mechanism models fail would settle the mechanism question for Cu and W.

Only after both checks do we apply the validated forward model to bonded-film broadening of the kind seen in Quadrupolar effects on nuclear spins of neutral arsenic donors in silicon. There, the question is whether interface roughness contributes measurably to the strain spread beyond uniform thermal mismatch.