Counting Hydrogen at Interfaces: Tritium Autoradiography as a Microstructure-Wide Association Study

Hydrogen makes strong steels crack early, and it does this by collecting at particular internal features such as grain boundaries, carbide interfaces and the boundaries between martensite laths. The most direct way to see where it collects is tritium autoradiography. In this technique, radioactive hydrogen (tritium) is loaded into a thin metal foil and held at liquid-nitrogen temperature so it stays put. It is then covered with a photographic emulsion, and each decay leaves a silver grain that shows up in the same electron micrograph as the crystal defects beneath it. For thirty years this technique has produced pictures and verbal verdicts, such as "tritium sits on incoherent carbide interfaces, not coherent ones." It has not produced numbers. Meanwhile, embrittlement models such as the Huang et al. fracture-toughness formula need one number above all: the fraction of interface trap sites occupied by hydrogen.

We propose to turn autoradiographs into that number using the study design that population biobanks such as UK Biobank made routine. Each interface segment in a micrograph is treated as an "individual." Its structural attributes (type, misorientation, coherency, neighbouring precipitates) play the role of the genotype, and its silver-grain yield plays the role of the phenotype. Nuisance variables such as foil thickness and emulsion batch are adjusted for the way genetic studies adjust for population structure, and every finding must replicate on specimens that were not used to discover it. Put in two sentences: we fit a physical measurement model that assigns each silver grain probabilistically to the traps that could have produced it, and we estimate per-interface hydrogen occupancy across thousands of segments with biobank-grade statistical discipline. The test is whether these occupancies explain which heat treatment embrittles a steel most, where grain size and bulk permeation measurements do not.

Core Claim

Occupancy per unit interface, estimated this way, is a transferable and predictive quantity. Multiplied by the measured interface density of each microstructure, it should predict the non-monotonic embrittlement of quenched S355 steel. Yan et al. report that the embrittlement index peaks at 61.06% for 1050 °C quenching, even though grain size keeps growing from 29.61 µm (950 °C) to 96.52 µm (1100 °C). Occupancy should also supply the coverage term in the Huang et al. toughness model from direct observation rather than from a fitted isotherm. No step in the existing literature produces such a number. What is new is the measurement model together with the association-study design. Neither the tritium technique nor the steels is new.

How the Three Directions Enter

Tritium autoradiography is the measurement instrument. The key method papers are Mori et al., who showed that 77 K exposure stops tritium from moving during exposure, and Miyazawa et al. and Kasuya, who established the incoherent-versus-coherent interface contrast in SUS316L. These papers motivate the proposal because they show that the signal is there. They also show that interpretation has stayed visual. None of them reports grain counts per interface class with uncertainty, and none corrects for grains that land between two nearby features. Aoki et al. paired autoradiography with thermal desorption spectroscopy, which measures bulk hydrogen release during heating. We reuse that pairing as a mass-balance constraint.

Hydrogen embrittlement is the application and the place where predictions are tested. Yan et al. and Wu et al. used the room-temperature hydrogen microprint technique together with permeation and slow-strain tensile tests. They found that lath martensite and lath bainite boundaries drive susceptibility. Wu et al. report an embrittlement index of 21.9% for water cooling, 15.45% for air cooling and 10.98% for furnace cooling in X80 steel. Both papers supply the target phenomena, and their open problem motivates our work. Yan et al. could not separate the contribution of prior-austenite grain size from that of lath-boundary density. Huang et al. supplies the model we feed:

$$K_{IH} = K_{IC}\,(\beta' K_{IC})^{(1-\eta\,\theta_H)},$$

where $K_{IH}$ is the toughness with hydrogen present, $K_{IC}$ the intrinsic toughness, $\beta'$ a material constant, $\eta$ a coverage-sensitivity parameter and $\theta_H$ the interface trap coverage. Huang et al. obtain $\theta_H$ from a Langmuir–McLean isotherm. We would measure it.

UK Biobank enters only as a methodological lens. It contributes no data, and we do not pursue the implant-failure angle, because UK Biobank records do not identify implant alloys or failure mechanisms. We borrow three habits. The first is exposure-wide screening with multiple-testing control, as in the 369-variable Bonferroni screen of Felici et al.. The second is external replication in a separate cohort, as Sigurdsson et al. did with the Danish Blood Donor Study. The third is the question of whether effect sizes carry over between groups, which Hu et al. asked across human ancestries. We ask it across alloys: does a lath boundary in S355 trap like a lath boundary in X80?

Research Program

The core object is a spatial point-process model of silver grains, meaning a model of where random points are likely to fall. Let $\lambda(x)$ be the expected density of silver grains at image position $x$. We model it as

$$\lambda(x) = b_s + \epsilon_s \sum_k \theta_{c(k)}\, n_{c(k)} \int_{\Gamma_k} G_{t_s}(x-y)\,dy .$$

Here $b_s$ is the background fog of specimen $s$, and $\epsilon_s$ is its development efficiency. The sum runs over imaged features $\Gamma_k$, which are interface segments, dislocation lines and precipitate surfaces. Each feature has class $c(k)$, with trap-site density $n_c$ per unit length and occupancy $\theta_c$. $G_{t_s}$ is a blur kernel set by the beta-particle range and by emulsion crystal size, and it widens with foil thickness $t_s$.

Only the product $\theta_c n_c$ can be identified from images alone. Two external measurements separate the terms and set the absolute scale. Thermal desorption peak areas on sister specimens give total hydrogen per trap population, and permeation trap densities of the kind Yan et al. report give $n_c$. Inference is Bayesian. Each grain gets a posterior probability of having come from each nearby feature or from fog, so a grain between two laths 150 nm apart is split between them rather than given to the nearer one.

On top of this model sits the association layer. Every segment is annotated with its class and continuous attributes: misorientation, coherency (from diffraction), adjacent carbide type and distance to triple junctions. Segment-level occupancy is regressed on these attributes. Foil thickness, exposure time, emulsion batch and position relative to the foil edge enter as covariates, just as ancestry principal components enter a genetic association study. Discovery and replication use disjoint specimens prepared on different dates.

The specimens are S355 quenched from 950, 1000, 1050 and 1100 °C, following Yan et al., and X80 cooled three ways, following Wu et al. SUS316L aged to form $\text{Cr}_{23}\text{C}_6$ serves as the reference system where the qualitative answer is already known.

One caveat must be stated in advance. Tritium and ordinary hydrogen differ in mass, so their trapping energies and mobilities may differ slightly. We treat this as a possible systematic offset. We check it through the thermal desorption peak positions and do not assume it away.

Baselines

The do-nothing reference is the current practice in Miyazawa et al. and Kasuya: expert visual classification of which features carry silver. We run it by having two metallurgists label the same micrographs blind. It is the natural first comparison because it is how every existing conclusion was reached. Beating it means producing occupancy estimates with calibrated intervals that separate interface classes the experts rate as equivalent, or that overturn an expert call and then replicate.

The naive quantitative baseline is stereological counting. Each grain is assigned to its nearest feature, and grains per unit length are compared across classes. This is the strongest simple competitor. It uses the same micrographs, feature tracings and segment annotations as our model, differing only in how grains are attributed. We beat it if, on synthetic ground truth, it shows bias that grows as lath spacing approaches the blur width while our model does not. On real data, the two methods must then disagree in the densest lath-martensite conditions, since that is where attribution matters.

The strongest existing route from microstructure to embrittlement is the microprint-plus-permeation analysis of Yan et al.. It explains the embrittlement index through trap density and lath-martensite fraction, and reports $I_z$ = 61.06% at 1050 °C with $D_{eff}$ = 2.08 × 10⁻⁶ cm²/s. We rerun the permeation and tensile measurements on our own S355 batches so both predictors face the same targets. The comparison is a leave-one-condition-out prediction of the embrittlement index. Beating Yan et al. requires correctly ordering all four quenching temperatures, including the peak at 1050 °C, where bulk trap density alone does not. For the fracture side, the baseline is the Huang et al. model with its fitted isotherm. We compare it against the same model with measured $\theta_H$, holding $\eta$ and $\beta'$ fixed. No reported result was found on Lacuna for the numerical fitting error of Huang et al., so we recompute it ourselves on our data.

Two ablations isolate the key ingredients. The first removes the blur kernel, which reduces the model to soft nearest-feature counting. The second removes the specimen covariates, which tests whether foil thickness and batch really act as confounders. No reported result was found on Lacuna for the accuracy of any quantitative grain-to-trap assignment in autoradiography, so the synthetic benchmark we build is itself a contribution.

Why not just…?

Why not just use the hydrogen microprint technique, which needs no radioactivity? The microprint technique records hydrogen as it leaves the sample at room temperature. It therefore shows the paths hydrogen escapes along, not where it sits. Mori et al. showed that room-temperature mobility blurs the link between a hydrogen atom and its trap. Its resolution also cannot resolve individual laths or tell coherent from incoherent carbide interfaces, and that contrast is exactly what Miyazawa et al. needed tritium to see. It remains a real alternative for ranking microstructures, and our strongest baseline is built on it. If it predicts the S355 ordering as well as our occupancies do, the added cost of tritium is not justified for that purpose, and we will say so.

Why not just deconvolve thermal desorption spectra into trap populations? Desorption spectroscopy gives binding energies and amounts but no locations. When lath boundaries and dislocations have overlapping binding energies, as in the deformation structures studied by Aoki et al., the spectrum cannot split them. We use it to set the absolute scale of the model, not to identify traps. The decisive experiment is to take two S355 conditions with similar desorption spectra but different embrittlement and check whether their per-class occupancies differ.

Why not just count grains next to boundaries? In ferritic–pearlitic X80 or aged 316L, feature spacing is much larger than the blur width, and plain counting will probably match our model. We expect it to. The model earns its place only in lath structures, where neighbouring traps lie within the blur distance. The synthetic benchmark and the counting-versus-model comparison on water-cooled X80 and 1050 °C S355 settle this directly. If the two agree there too, the simpler method wins.

Why not just train a deep image model, such as the locally connected networks of Sigurdsson et al., from micrographs to embrittlement? There are only a handful of heat-treatment conditions, so such a model would have seven labels and thousands of parameters. It would also return a prediction without an occupancy value that can feed a physical toughness model. Deep networks may still help with the tedious part, tracing boundaries in micrographs, and we will use a segmentation model there. The inference itself stays explicit.

Proposed Experiments and Decisive Checks

The first check is synthetic. We place grains according to known occupancies on real traced S355 and X80 feature maps, varying blur width, fog and lath spacing. We then report bias and interval coverage for the full model, nearest-feature counting and the two ablations. The model should keep 90% intervals at nominal coverage where counting drifts.

The second check is recovery of known physics on SUS316L. The model must reproduce near-zero occupancy on coherent interfaces and clearly positive occupancy on incoherent $\text{Cr}_{23}\text{C}_6$ interfaces. It must do so in the replication set, not only the discovery set. Failure here would mean the measurement model is wrong.

The third check is the main test. For each S355 condition, we multiply lath-boundary and prior-austenite-boundary occupancies by their measured densities. We ask whether this reproduces the peak at 1050 °C and its decline at 1100 °C, and whether it apportions the embrittlement between the two boundary types, which Yan et al. could not do. We then plug the measured coverage into the Huang et al. model and compare the predicted $K_{IH}$ with fracture tests on the same batches.

The fourth check is transfer. Lath-boundary occupancies estimated in S355 are used, without refitting, to predict the X80 ranking of water cooling above air cooling above furnace cooling. This follows Hu et al.'s question of whether effect sizes are conserved between groups. Conserved occupancies would mean an interface class carries a portable trapping coefficient, and any mismatch would show which boundary attributes matter beyond class.