Grain Size as a Bottleneck: Hall-Petch-Structured Learning from a Dozen Laser Tracks

When engineers tune a laser deposition process to make a steel harder, they are mostly tuning one hidden quantity: how small the crystal grains end up. Smaller grains block the defects that carry permanent deformation, so the metal resists yielding and indentation. Physicists summarize this in the Hall-Petch law, where strength rises with the inverse square root of grain size. Yet the machine-learning models used to pick process settings from small experimental campaigns ignore that hidden quantity. They map laser power, scan speed and pulse shape straight to hardness, and pad their nine to fifteen data points with synthetic noise. We propose the opposite design. The model must route every prediction through grain size, a variable it is told exists, measured on only a fraction of specimens, and linked to hardness by the Hall-Petch form. The parameters of that form are pooled across alloys, so a new alloy needs only a handful of runs to calibrate.

The claim is specific and testable. With 10–15 deposition runs per alloy and electron backscatter diffraction (EBSD, an expensive electron-microscope method that maps crystal orientation and therefore grain boundaries) on roughly a third of them, a partially observed grain-size bottleneck should predict hardness better than flat regressors. It should also adapt to a held-out alloy from about five runs, where flat models need many more. If a hand-fitted two-step physics model does just as well, we will report that plainly. That outcome would mean the physics structure, not the learning, carries the value.

Why this gap exists

Both of the paper's two areas supply half the story. On the materials side, rapid-solidification studies show that process settings act on hardness largely through refinement. In Influences of Pulse Shaping on Single-Track Clad of AISI316L Stainless Steel by Laser Material Deposition, changing only the laser's on-off timing at fixed 1400 W peak power changes cooling rate, grain size and microhardness together. Rectangular pulses give the finest grains and the highest hardness, while continuous wave gives the coarsest and softest. Comparison on Microstructure and Properties of Stainless Steel Layer Formed by Extreme High-Speed and Conventional Laser Melting Deposition shows refinement compensating for a softer retained-austenite phase, keeping AISI 431 at 600–720 HV. Nano-mechanical Behavior of H13 Tool Steel Fabricated by a Selective Laser Melting Method links a 1.42 µm granular structure to 8.41–9.18 GPa nanohardness, against about 6.6 GPa for conventional H13. These papers motivate the bottleneck: the causal chain runs from process settings to cooling rate, then to grain size, then to hardness. Each paper, however, characterizes one alloy, a few settings and no predictive model.

On the learning side, the clearest evidence that structure substitutes for data comes from clinical records. MiME: Multilevel Medical Embedding of Electronic Health Records for Predictive Healthcare builds the visit-diagnosis-treatment hierarchy into its model. It reports its largest gain, a 15% relative improvement in PR-AUC, on the smallest cohort. GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs attaches outside knowledge graphs and matches, with 36 patients, baselines trained on 720. These papers supply the principle we reuse: encode the known intermediate structure of the problem, then let the data fill in the constants. They do not supply a method we can transplant, because a medical code hierarchy has nothing to do with a dislocation-blocking law.

The closest materials-ML work is Data-Driven Modeling of Dislocation Mobility from Atomistics using Physics-Informed Machine Learning. It splits a graph neural network into sub-networks mirroring an Arrhenius mobility law, and uses ensemble-uncertainty active learning to cut molecular-dynamics runs by 40%. We share its stance that physics should shape the network, but differ in level and data source. It models a single defect in pure single-crystal iron from simulations, and it explicitly leaves out grain-boundary interactions and Hall-Petch strengthening. We model polycrystalline strengthening at the process level from physical experiments, which is exactly what it leaves out.

The small-data manufacturing model we most directly oppose is Machine Learning Prediction of Turning Precision Using Optimized XGBoost Model. It fits nine Taguchi runs with SMOGN oversampling (synthetic points made by adding Gaussian noise between existing ones) plus swarm-optimized hyperparameters. Finally, Generalizing Process Engineering Models Across Material Families documents why such models break across alloys: they interpolate within one grade and carry no physical invariants. That finding constrains our design. Whatever transfers across alloys must be physical structure, not fitted correlations.

Core Claim

Let $x$ be a run's process settings together with a cheap, physics-derived estimate of its cooling rate $\dot T$ (an analytic moving-heat-source calculation from power, speed and pulse timing). Let $a$ index the alloy. The model has two linked parts:

$$\log d = \log A_a - n \log \dot T + g_\phi(x) + \varepsilon_d$$

$$H = H_{0,a} + k_a\, d^{-1/2} + \beta_a f_{\text{LAGB}} + \varepsilon_H$$

The first equation predicts grain or cell size $d$. Here $A_a$ and the exponent $n$ come from the standard solidification-spacing power law, $g_\phi$ is a small Gaussian-process correction for effects the cooling-rate estimate misses, and $\varepsilon_d$ is noise. The second equation predicts hardness $H$. Here $H_{0,a}$ is the alloy's lattice hardness, $k_a$ its Hall-Petch slope, and $f_{\text{LAGB}}$ the measured or predicted fraction of low-angle boundaries (boundaries misoriented by less than 15°) with coefficient $\beta_a$.

Grain size $d$ is observed on the EBSD subset and treated as a latent variable elsewhere. Hardness is observed everywhere. The exponents and the correction $g_\phi$ are shared across alloys. The alloy-specific constants $(A_a, H_{0,a}, k_a, \beta_a)$ get hierarchical priors, so a new alloy borrows strength from the others. Inference is fully Bayesian, which yields calibrated intervals; with 15 points per alloy, point estimates alone would mislead.

The low-angle term addresses a paper that cuts against the plain Hall-Petch story. Unveiling the Deformation Mechanism of High-Strength Low-Alloy Structural Steel with Gradient Dislocation-Cell Structure raised yield strength from 371.5 to 522.3 MPa through cells walled mostly by low-angle boundaries and dense dislocation walls (below 2°), plus stresses from deformation gradients. A single-slope Hall-Petch law, which assumes strong high-angle barriers, cannot describe that. We therefore expect the bottleneck to help most where boundaries are high-angle and least where low-angle cells dominate, and we will test that prediction directly.

Research Program

The work has three parts. First, we build a dataset, because nothing on Lacuna serves this question. We deposit single-track and small multi-track clads in four steels whose refinement behavior the file documents: 316L, AISI 431, H13, and a low-alloy structural steel. Each alloy gets 15 runs spanning power, scan speed and the five pulse shapes of the 316L study. All specimens receive Vickers microhardness traverses. We use microhardness rather than nanoindentation because the H13 paper itself warns that nanoscale hardness does not translate to macroscopic properties. About five specimens per alloy receive EBSD, giving grain size and low-angle fraction. We will release the data, including which specimens were deliberately left unmeasured.

Second, we fit the model and its ablations under identical leave-one-out and leave-one-alloy-out protocols. Third, we close the loop as the turning-precision paper did. The fitted model recommends settings for a target hardness on the held-out alloy, and we deposit and measure them.

A secondary question is whether the model should choose which specimens get EBSD. We will compare random EBSD allocation against allocation by posterior uncertainty in $d$, in the spirit of the uncertainty-driven querying in the dislocation-mobility paper. We will claim a benefit only if it exceeds the spread across random allocations.

Baselines

The naive reference point is a per-alloy linear regression of hardness on raw process settings, alongside the alloy-mean predictor. Any method that cannot beat these under leave-one-out on 15 points has learned nothing.

The strongest existing small-sample manufacturing pipeline is the SMOGN-plus-swarm-optimized XGBoost of Machine Learning Prediction of Turning Precision Using Optimized XGBoost Model. It is the most direct competitor because it targets exactly the nine-to-fifteen-run regime with off-the-shelf tools. On its nine-run machining dataset it reports a leave-one-out R² of 0.661 (RMSE 0.006878), up from 0.554 for plain XGBoost. That number comes from a different response, machining error rather than hardness, so it is not a target on our data. It shows what this pipeline achieves when its authors tune it. We will reimplement the pipeline and, if the nine runs can be recovered from that paper, first check that we reproduce roughly 0.661 before running it on our data. It then gets the same process features, the same cooling-rate estimate as an input, the same folds, and a hyperparameter search budget equal to the compute we spend on our posterior inference.

A Gaussian-process regressor on process settings plus cooling rate is a further flat baseline. It is often strongest at this sample size and needs no oversampling.

For cross-alloy transfer, the competitor is a single pooled flat model, a Gaussian process with alloy descriptors such as thermal conductivity and composition, trained on all source alloys. We also run instance reweighting of source-alloy runs toward the target alloy, the idea behind the transfer survival boosting of Bellot et al..

Three ablations isolate the key ingredient. The first removes the EBSD labels, leaving $d$ fully latent. The second fits each alloy independently, removing cross-alloy pooling. The third replaces the Hall-Petch link with a free monotone function of $d$.

Beating the field means three things. Within alloys, hardness RMSE must be lower than the best flat baseline across leave-one-out folds, with the difference exceeding fold-to-fold variability. For a held-out alloy, the learning curve must reach the flat models' 15-run error with no more than about five target runs. Prediction intervals must cover near their nominal rate.

Why not just…?

Why not just fit a flat model with oversampling or a Gaussian process? The turning-precision study shows that oversampling plus heavy tuning still left about a third of variance unexplained on nine runs. The cross-material analysis argues that such models interpolate within one grade and fail on the next. A Gaussian process may nonetheless be competitive within a single alloy, and we expect that. The decisive comparison is the held-out-alloy learning curve, where a flat model has no shared structure to carry over. If it matches us there too, the bottleneck is not earning its keep.

Why not just fit Hall-Petch and a cooling-rate law by hand, in two steps? This is the real threat. A materials scientist would fit the power law on the EBSD specimens, fit Hall-Petch on the same specimens, and chain them. That uses only the five measured specimens per alloy to fit hardness, and it shares nothing across alloys. Our model also uses hardness on the ten unmeasured specimens to inform grain size, and pools exponents across alloys. The independent-alloy and no-EBSD-label ablations, together with an explicit two-step baseline, settle the question. If the two-step fit is within noise of the full model, we will report that the physics form, not joint inference, is the contribution, and recommend the simpler tool.

Why not just generate more training data with a generative model? Diffusion synthesis helped on small EEG cohorts in DiffMDD: A Diffusion-Based Deep Learning Framework for MDD Diagnosis Using EEG. However, Assessing the Potential of Data Augmentation in EEG Functional Connectivity for Early Detection of Alzheimer's Disease found gains capped near doubling the data and losses for models sensitive to fine structure. Synthesizing process-hardness pairs from 15 points can only echo whatever relationship the generator already assumes. We include SMOGN as the representative augmentation and do not expect generative synthesis to add information the bottleneck lacks.

Proposed Experiments

The first decisive check is within-alloy leave-one-out hardness error for all methods across the four steels. We break results down by low-angle boundary fraction to test the prediction that gains shrink in cell-dominated microstructures such as the low-alloy steel. The second is leave-one-alloy-out transfer, plotting held-out-alloy error against 1, 3, 5, 10 and 15 target runs, together with interval coverage. The third is grain-size prediction on specimens the model never saw measured, scored against EBSD held back for evaluation. This checks whether the latent variable is physically meaningful rather than a free fudge factor. The fourth is the confirmation deposition: settings recommended for a target hardness on the held-out alloy, deposited, and measured against the predicted interval.

The proposal fails cleanly in two cases: if the hand-fitted two-step model matches the full model in transfer, or if inferred grain sizes disagree with held-back EBSD while hardness predictions still look good. Either result is publishable, because it tells practitioners which part of the physics to bother encoding.