Metallurgists want to predict, before running an experiment, whether a new alloy will deform by twinning, meaning a block of crystal shears into a mirror-image orientation. Twinning is the mechanism behind the unusual combination of strength and ductility in many high-entropy alloys. The underlying physics is well characterized. Whether a face-centered cubic metal twins depends on three energy barriers along a sliding path: the stacking fault energy, the unstable stacking energy, and the unstable twinning energy. Data-driven predictors of these quantities, however, are fit mostly on pure elements and simple binary alloys, and they are then applied to concentrated multi-element alloys whose chemistry they never saw. Statistical genetics has studied this failure pattern for a decade under the name of ancestry bias. Risk models trained on European genomes lose accuracy in other populations because the measured markers are only proxies for the causal variants, and the correlation between proxy and cause differs from one population to the next. The most effective fix found so far is to measure portability locally, region by region, rather than globally.
This note proposes treating alloy families as "populations" and running that playbook on twinnability. The core idea fits in two sentences. Average composition plays the role of a proxy marker. The causal variable is the local chemistry of the atoms sitting on the slipping plane, and the link between the two shifts across alloy families and with chemical short-range order, much as proxy-to-cause correlation shifts across human ancestries. A predictor that models the local fault-plane chemistry, and that uses a locally estimated portability score to decide how far to trust what it learned on simple alloys, should transfer twinnability rankings to new alloy families from far less target data than global-composition models or naive fine-tuning need.
Core Claim
Two prior findings share the same structure, and this note builds on that shared structure. In metallurgy, Tadmor and Bernstein showed that the single global stacking fault energy cannot explain why aluminum almost never twins while palladium twins readily. The ordering Pt < Al < Ir < Au < Cu < Pd < Pb < Ag only appears once the barrier landscape is decomposed into the unstable stacking and unstable twinning energies. In genetics, Miao et al.'s X-Wing showed that cross-ancestry prediction improves when shrinkage is relaxed only in genomic windows with high local genetic correlation. For mean corpuscular volume, windows covering about 1% of the genome carried over 60% of the European–East Asian genetic covariance. Both papers motivate this proposal. Neither tests whether local portability weighting helps transfer physical barrier energies across chemical backgrounds.
The specific, falsifiable claims are three. First, transfer error for barrier-energy predictors is concentrated in a small set of local fault-plane chemical environments, mirroring the X-Wing localization result. Second, when two alloys share the same average composition but differ in short-range order, composition-keyed predictors fail by construction. This is the materials analogue of linkage-disequilibrium heterogeneity. Local-environment predictors should not fail in that setting. Third, portability-weighted transfer recovers twinnability rankings in held-out alloy families with tens rather than hundreds of target calculations.
Research Program
The first component is a dataset, because the file contains no large public collection of generalized stacking-fault curves for concentrated alloys. We will compute stacking fault energy, unstable stacking energy, and unstable twinning energy with density functional theory on supercells containing explicit fault planes. Each composition gets many random chemical realizations, so each one yields a distribution of local fault-plane environments and energies rather than a single average.
The calculations fall into three groups. The source population is pure FCC elements plus Ni-, Cu-, Co-, Fe- and Cr-based binaries. Target population A is Cantor-based alloys with roughly 5 at.% refractory additions (Nb, Ta, Mo, W), following the compositions down-selected in the refractory-chemistry MPEA study. That study used KKR-CPA and Ising-type models to place useful twinning/transformation behavior at stacking fault energies of 15–30 mJ/m². Target population B is equiatomic NiCoCr, studied by Banerjee et al., with short-range order generated by Monte Carlo annealing at several temperatures. The NiCoCr group gives identical average compositions with different local chemistry. The elemental group doubles as a sanity anchor: the pipeline must reproduce the Tadmor/Bernstein ranking.
The second component is the predictor. Each fault configuration is described by counts of chemical species in shells around the slipping plane, with effects $\beta_k$ assigned to local environment classes $k$. A cross-family local correlation scan estimates $\rho_k$, the correlation between source-fitted and target-fitted effects within environment class $k$. This is computed from the small target sample by resampling, in the spirit of X-Wing's window scan. The target effects then receive a prior centered on the source effects, with confidence set by local portability:
$$ \beta_k^{\text{tgt}} \sim \mathcal{N}\!\left(\rho_k\,\beta_k^{\text{src}},\ \tau^2(1-\rho_k^2)\right), $$
Here $\tau^2$ is a global prior variance fit by cross-validation. Where $\rho_k$ is high, the target model borrows from the source. Where it is low, the target data decide. The form is the standard conditional distribution of one correlated Gaussian effect given another.
Predicted barrier energies feed into Bernstein's closed-form twinnability approximation, which is reported to be within 0.2% of the full criterion. This keeps the endpoint physically interpretable. The usual caveat applies: the criterion is a zero-temperature, quasistatic quantity.
Baselines
The naive reference point is the source model applied unchanged to the target family. A second naive reference is the classical rule that low stacking fault energy implies twinning, applied together with the 15–30 mJ/m² window from the MPEA study. Both baselines get the same density functional theory test sets. Beating the stacking-fault-energy rule on twinnability ranking within target families is the minimum bar. It is plausible because the Al/Pd anomaly already shows that rule failing on elements.
The strongest learned competitor is a periodic graph transformer, Matformer, run from its official code. It will be pretrained on the source population and fine-tuned on each target budget, with the same number of hyperparameter trials as our method and early stopping on a shared validation split. It is the strongest competitor because it sees full atomic structure, so in principle it could learn local fault-plane chemistry without being told to. We found no reported result on Lacuna for Matformer or any other learned model predicting unstable stacking or unstable twinning energies in concentrated FCC alloys. The only reported Matformer splits in the file are for formation energy, band gap, and elastic moduli. A second learned baseline is pooled source-plus-target training of the same local-environment model with uniform weights, the analogue of naive multi-ancestry meta-analysis.
Two ablations isolate the key ingredients. The first keeps portability weighting but uses global-composition features, which tests whether local chemistry matters. The second keeps local-environment features but sets $\rho_k$ to a single global value, which is the analogue of genome-wide shrinkage of the kind X-Wing improved upon and tests whether locality of the transfer weight matters.
Winning means lower target mean absolute error on all three barrier energies, and higher Spearman correlation of predicted twinnability rankings across held-out compositions. Both must hold at target budgets of 10 and 50 configurations, with confidence intervals from resampling compositions rather than configurations.
Why not just…?
Why not just fine-tune Matformer on the target alloys? This is the real threat, and it is tested directly. Our expectation is that at very small target budgets a large graph model either barely moves from the source fit or overfits. The explicit $\rho_k$ prior gives a principled middle ground and shows where the source knowledge stops holding. If fine-tuning matches us at 10 and 50 configurations, the predictive claim fails. The diagnostic claims, localization of transfer error and short-range-order sensitivity, would still stand on their own.
Why not just fit a local-environment model on target data alone and skip transfer? If target calculations are cheap, transfer is pointless. The learning-curve experiment settles this: we report the target budget at which target-only training catches up. The method is worth using only if that crossover lies well above the budgets affordable when screening the roughly one million candidate compositions considered in the MPEA screening study.
Why not just compute the three barriers with density functional theory for every candidate? That calculation is the ground truth, and it is what we use to build the test sets. At screening scale it is unaffordable. The cheaper tight-binding route deviates from experiment by roughly 22.8–52.2% on stacking energies (Tight-binding calculations of stacking energies and twinnability in fcc metals), and empirical potentials struggle with stacking energies in NiCoCr according to Banerjee et al.. Transfer learning is therefore competing with these approximations, not with full density functional theory.
Why not just select invariant features, as in Invariant Ancestry Search? That method finds predictor subsets whose relationship with the outcome is stable across environments. Applied to local-environment features, it is a sensible extra baseline and will be run. It makes a hard keep-or-drop decision per feature, whereas the portability prior partially borrows strength. The comparison shows whether that soft middle ground matters.
Decisive Checks
The short-range-order test comes first. For NiCoCr samples with identical composition but different annealing temperature, we measure the spread of density functional theory barrier energies. If the spread is negligible, the linkage-disequilibrium analogy is empty, composition-only models are sufficient, and the proposal shrinks to an ordinary transfer-learning study. If the spread is large, composition-keyed predictors must show an error floor that local-environment predictors do not.
Second, we test localization: what fraction of local environment classes accounts for most of the source-to-target residual variance. This is reported the way X-Wing reports genomic coverage against covariance explained.
Third, we test calibration of the scan: environment classes flagged with low $\rho_k$ on one target split should carry larger residuals on an independent split.
Fourth, the elemental sanity check: the pipeline must recover the Pt < Al < Ir < Au < Cu < Pd < Pb < Ag ordering.
A positive outcome on all four would give materials screening a quantitative statement of where its models stop transferring. The epidemiology literature has spent years building that kind of statement, and alloy design currently lacks one.
References
- A first-principles measure for the twinnability of FCC metals
- Tight-binding calculations of stacking energies and twinnability in fcc metals
- Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics
- Understanding the effect of refractory metal chemistry on the stacking fault energy and mechanical property of Cantor-based multi-principal element alloys
- An atomistic study connecting underlying dislocation behavior with superior mechanical properties of NiCoCr medium entropy alloy
- Periodic Graph Transformers for Crystal Material Property Prediction
- Invariant Ancestry Search