Hydrogen weakens steel, and it does so unevenly. It collects at particular internal interfaces, and those interfaces are where cracks begin. The most direct way to see where hydrogen sits inside a steel is tritium autoradiography. Radioactive hydrogen (tritium) replaces ordinary hydrogen. Each decay fires a short-range electron into a photographic emulsion coated on a thin foil. After development, the emulsion leaves silver particles that can be overlaid on a transmission electron microscope image of the same foil. The results are striking. Coherent twin boundaries hold essentially no hydrogen, while disordered grain boundaries, twin steps and the mismatched faces of carbide precipitates hold a great deal. The trouble is that each paper reports its numbers under its own protocol: a different exposure temperature, a different wait before exposure, different emulsion grain sizes, even different units. A count of "15 particles" in one paper cannot be set beside "15 particles" in another.
We propose to treat that problem the way a careful meta-analyst treats studies that measured the same thing with different scales. Concretely, we will write down a physical model of how a true hydrogen distribution becomes a pattern of silver dots under a given protocol. We will fit it jointly to autoradiographs from a deliberately crossed set of protocols on sister foils. The model reports trap strength per interface type in units that do not depend on the protocol. The same design also lets us separate two effects that current work cannot tell apart: disorder in the grain boundary itself, and carbides sitting on that boundary.
Core Claim
The disagreements among published tritium autoradiography trap densities are mostly measurement effects that can be modelled, not differences in the traps. Three effects matter. The first is spatial blur, set by exposure temperature and silver-grain size. The second is loss of weakly bound hydrogen during the hold before exposure. The third is detection efficiency, set by emulsion and development. Once these are modelled explicitly, one set of interface-class trap intensities should explain autoradiographs taken at 253 K and 77 K, and with and without a room-temperature hold. That set of intensities should also predict the count pattern under a protocol held out of the fit.
The model treats the silver particles in a field of view as a random scatter of points, a spatial point process. Its expected density at position $x$ under protocol $p$ is
$$\lambda_p(x) = \eta_p \int K_p(x-y)\, \sum_{c} \theta_c\, s_{c,p}\, \mathbf{1}_c(y)\, dy + \beta_p .$$
The symbols are as follows. $\theta_c$ is the protocol-free tritium occupancy per unit interface area for interface class $c$ (coherent twin, twin step, low-index grain boundary, high-index grain boundary, coherent carbide face, incoherent carbide face). $\mathbf{1}_c(y)$ marks where class $c$ lies in the microstructure image. $K_p$ is a blur kernel whose width includes electron range, silver-grain size and diffusion during exposure. $s_{c,p}$ is the fraction of class-$c$ hydrogen that survives the protocol's pre-exposure hold. $\eta_p$ is emulsion detection efficiency. $\beta_p$ is the uniform background from lattice hydrogen. The survival fraction is the only class-and-protocol-dependent term. We model it as $s_{c,p}=\exp(-t_p\,\nu\, e^{-E_c/k_BT})$, where $t_p$ is the hold time, $\nu$ is an attempt frequency, $E_c$ is the trap binding energy, $k_B$ is Boltzmann's constant and $T$ is the hold temperature. This is ordinary first-order detrapping kinetics. The model is new as applied to this measurement. The equation itself is standard.
What Prior Work Establishes and Leaves Open
Four tritium autoradiography studies on SUS316 and SUS316L supply the phenomenon and the numbers our model must reproduce. Observation of Hydrogen Trapping Sites at 77 K Exposure in SUS316L Steel by means of Tritium TEM Autoradiography exposes foils in liquid nitrogen for 90 days. It reports about 0.04 µm resolution against about 0.12 µm at 253 K. It also gives a ranking of about 35, 15, 10 and 0 particles per 0.11 µm² for Cr₂₃C₆ carbides, grain boundaries, twin steps and coherent twins. Observation of Hydrogen Trapping Sites in Long-Term Aged Stainless Steel by Tritium TEM Autoradiography compares immediate exposure with a 3-day hold. It finds 15–17 against 2–3 particles per 0.1 µm², a roughly sevenfold ratio. By tilting the specimen, it shows that only the incoherent faces of a single carbide trap hydrogen. Observation of Hydrogen Trapping in Grain Boundaries and Grain Boundary Precipitates by Tritium TEM Autoradiography ties heavier trapping to boundaries on high-index planes and reports trapping per length: about 32 particles along 3.23 µm. Tritium Transmission Electron Microscopic Autoradiography of Hydrogen Trapping Sites at Interfaces in an Austenitic Stainless Steel SUS316L uses 1000 kV microscopy and finds near one-to-one alignment of silver with boundary dislocations. It names the resolution floor, about 0.23 µm developed silver from about 0.12 µm AgBr, and notes that carbides sitting on boundaries confound the attribution of trapping.
These papers motivate the proposal. Between them they report three different normalizations (per 0.11 µm², per 0.1 µm², per length), two exposure temperatures and two hold regimes. None of them models how these choices change the counts. Every density was obtained by manual counting in a few fields. We reuse their microscopy and charging methods unchanged. What we add is the measurement model and a crossed experimental design.
The heterogeneity lens comes from two sources. One is clinical synthesis work, where pooling fails because endpoints are defined differently, as with the incompatible bleeding scales in antiplatelet trials (Jiaxin Yang et al.). The other is computational materials science. Automated determination of grain boundary energy and potential-dependence using the OpenKIM framework runs one standardized test over 225 interatomic potentials. It shows that the qualitative energy landscape is shared while absolute magnitudes differ by a factor of 2 to 3. We borrow its logic: a standardized procedure across "instruments" separates what is invariant from what is instrument-specific. The difference is that our instruments are experimental protocols rather than potentials, and our quantity is hydrogen occupancy rather than dry boundary energy. Fundamental Microscopic Properties as Predictors of Large-Scale Quantities of Interest: Validation through Grain Boundary Energy Trends and Application of Machine Learning to Predict Grain Boundary Embrittlement in Metals by Combining Bonding-Breaking and Atomic Size Effects regress boundary energetics on descriptors. The second pools 142 DFT strengthening energies and reaches r² ≈ 0.889. Neither includes hydrogen or experimental tracer data. We treat them as a constraint: descriptor regression is not attempted here until protocol-invariant targets exist.
Research Program
The experiment is a crossed design on sister foils cut from one SUS316L plate. Half the plate is solution-annealed, with no carbides. Half is aged at 923 K following Observation of Hydrogen Trapping Sites in Long-Term Aged Stainless Steel by Tritium TEM Autoradiography, so that Cr₂₃C₆ forms on some boundaries. Each condition is charged identically, then split across exposure temperature (253 K or 77 K) and hold (none or 3 days at room temperature). That gives eight cells, each exposed for 90 days. Before the emulsion goes on, each foil is mapped by diffraction to assign boundary misorientation, plane indices, twin coherency and carbide-face coherency. This produces the class masks $\mathbf{1}_c$.
The annealed arm is the key to Gap 3. There, grain boundary disorder appears without carbides. In the aged arm, some boundaries of matched misorientation carry carbides and others do not. The additive class structure can therefore identify boundary and carbide contributions separately. Coherent twins act as a built-in negative control: their true $\theta$ should be zero, so any counts near them measure blur leakage.
On the analysis side, we will build a silver-particle detector for the micrographs, trained on hand-annotated fields and checked against a second annotator. The file contains no prior automated-counting result for this imagery, so agreement between annotators sets the target. We will then fit the model above by Bayesian inference. The blur widths $K_p$ will have informative priors taken from the reported 0.04 and 0.12 µm resolutions, and the fit will report posterior $\theta_c$ and $E_c$. As a secondary product, we will extract the published densities from the four papers and re-express them through the fitted model. This tests whether the legacy literature is internally consistent once protocol effects are removed.
Linking these maps to bulk permeation trap density or to the embrittlement index measured in Effect of Quenching Temperature on Microstructure and Hydrogen-Induced Cracking Susceptibility in S355 Steel is deliberately left out of scope. Those are ferritic and martensitic steels with different trap populations, and a transfer function fitted on two alloys would be curve-fitting.
Baselines
The do-nothing reference is to take each cell's raw particle density per interface class, counted in a fixed-width band around the interface, and compare cells directly, as the source papers do. The strongest reported figures here are the 77 K densities of about 35, 15, 10 and 0 particles per 0.11 µm² in Observation of Hydrogen Trapping Sites at 77 K Exposure in SUS316L Steel by means of Tritium TEM Autoradiography. Our own 77 K, no-hold cell must reproduce this ranking before anything else is trusted. We beat this reference if raw counts disagree across cells for the same class while the model's $\theta_c$ agree.
The simple normalization baseline divides each interface count by the matrix background count in the same field. This cancels detection efficiency and is what a practitioner would try first. It is run on the same detected particles and fields. The model beats it only if the ratio still shifts with exposure temperature or hold while $\theta_c$ does not. Such a shift is expected, because ratios cancel neither blur nor class-dependent desorption.
The standard heterogeneity baseline is a random-effects meta-regression. It fits per-cell class densities with exposure temperature and hold as moderators, which is the default tool of the clinical synthesis literature. We give it the same cell-level data and the same moderators and let it fit freely. It explains variance but has no physical kernel, so the comparison is on predicting a held-out cell. For none of these methods is there a reported result on Lacuna for held-out-protocol prediction of autoradiographic counts.
Two ablations isolate the key ingredients. Removing $K_p$, so that counts come from fixed bands, tests whether blur modelling matters. Setting $s_{c,p}=1$ tests whether class-dependent desorption matters. All variants share the same detector output, the same fields and the same sampler budget.
Why not just…?
Why not just standardize on the 77 K protocol and rerun everything in one lab? This is the real threat, and the crossed design is partly a concession to it. A single protocol gives comparable numbers within that protocol. It cannot, however, separate diffusing from deep trapping, which requires the hold contrast. It also cannot reuse the existing 253 K literature. If the ablations show that $K_p$ and $s_{c,p}$ barely matter, then standardization alone is enough, and we will say so.
Why not just normalize by the matrix background? It is included as a baseline. The sevenfold loss after a 3-day hold reported in Observation of Hydrogen Trapping Sites in Long-Term Aged Stainless Steel by Tritium TEM Autoradiography is unlikely to be the same for every class. A single ratio should therefore fail across hold conditions. The held-out-cell test will show whether it does.
Why not just use thermal desorption or permeation cells? These average over the whole specimen. The facet-level asymmetry of single carbides, and the zero signal at coherent twins, are invisible to them. They remain useful as a check: the model's $E_c$ values should sit near the desorption peaks of a companion desorption measurement on the same plate.
Why not just compute trap energies atomistically? Boundary energies alone already vary two- to threefold across potentials (Automated determination of grain boundary energy and potential-dependence using the OpenKIM framework). Machine-learned potentials also show much larger force errors on hydrogen atoms than on metal atoms (Benchmarking of machine learning interatomic potentials for reactive hydrogen dynamics at metal surfaces). Simulation needs protocol-invariant experimental targets to be judged against, and this work supplies them.
Decisive Checks
The main test is to hold out one of the eight cells, for example aged, 77 K, 3-day hold. We fit on the other seven and predict its per-class counts and its count profiles perpendicular to the interfaces. We score log-likelihood and absolute error on density against the meta-regression and ratio baselines. Three further checks accompany it. First, the variance between protocols in $\theta_c$ should collapse toward zero, while it stays large for raw counts. Second, the fitted coherent-twin $\theta$ should be consistent with zero, with residual counts accounted for by the blur tails. Third, the carbide-face term should stay non-zero in the aged arm after conditioning on boundary class. That establishes precipitate incoherency as a trap separate from boundary disorder.
If the held-out cell is predicted no better than by the meta-regression, the measurement-model claim fails. The durable output would then be the crossed dataset and the automated counter.
References
- Observation of Hydrogen Trapping Sites at 77 K Exposure in SUS316L Steel by means of Tritium TEM Autoradiography
- Observation of Hydrogen Trapping Sites in Long-Term Aged Stainless Steel by Tritium TEM Autoradiography
- Observation of Hydrogen Trapping in Grain Boundaries and Grain Boundary Precipitates by Tritium TEM Autoradiography
- Tritium Transmission Electron Microscopic Autoradiography of Hydrogen Trapping Sites at Interfaces in an Austenitic Stainless Steel SUS316L
- Automated determination of grain boundary energy and potential-dependence using the OpenKIM framework
- Fundamental microscopic properties as predictors of large-scale quantities of interest: Validation through grain boundary energy trends
- Application of Machine Learning to Predict Grain Boundary Embrittlement in Metals by Combining Bonding-Breaking and Atomic Size Effects
- Effect of Quenching Temperature on Microstructure and Hydrogen-Induced Cracking Susceptibility in S355 Steel
- Benchmarking of machine learning interatomic potentials for reactive hydrogen dynamics at metal surfaces