Skip to main content
Back to timeline
arXivSource publication:

High-entropy alloy strength prediction stops paying per test: ROSE-HEA matches a model trained on 185 records using 125

Synopsis

Tensile-strength data for high-entropy alloys are costly to obtain, and random sampling often spends the budget on similar compositions and processing routes. ROSE-HEA ranks candidate experiments by how much their possible strength outcomes would reduce worst-case predictive risk across a reference composition–processing domain, while weights on competing smoothness hypotheses update with each new response. Across 501 records and 30 matched runs, it reached a mean RMSE of 152.84 MPa with 125 observations, below EI-BO and UCB-BO at the same budget and below XGBoost, the strongest conventional regressor, trained on 185 observations.

Source-provided article image: Sample-efficient prediction of high-entropy alloy strength through outcome-aware robust sequential learning
arXiv

Interpretation

Ultimate tensile strength of high-entropy alloys can be predicted with a smaller training budget: ROSE-HEA reached a mean RMSE of 152.84 MPa with 125 training observations, the lowest among the sequential strategies compared. Comparable sequential design relied on candidate-wise scores such as expected improvement or an upper confidence bound, which do not explicitly quantify how one observation improves predictions across a whole reference domain; random sampling can repeatedly characterize similar conditions. On 501 high-entropy alloy records across 30 matched runs sharing 25 initial observations and the same test set, ROSE-HEA's mean RMSE was 6.16% lower than EI-BO, 6.37% lower than UCB-BO and 11.89% lower than random GP sampling.

The selected observations carry transferable information rather than serving only ROSE-HEA's own predictor. If the gain came only from a particular model, the advantage of the training set should disappear when the regressor changes. Six regressor families fitted separately to ROSE-HEA-selected and randomly selected training sets all showed lower mean RMSE, with MLP falling from 202.88 to 174.89 MPa and XGBoost from 186.07 to 168.67 MPa.

In the composition–processing space, ROSE-HEA's selections spread more widely and bring larger prediction gains at distant conditions that the initial data leave sparse. Random sampling draws records with equal probability, so densely recorded regions are more likely to be selected repeatedly. In the full 14-dimensional input space, mean nearest-training distance fell from 0.8255 to 0.6658 and the 95th-percentile distance from 2.9166 to 2.3648; split into thirds by distance from the 25 initial observations, the distant group's mean RMSE dropped 18.95%, from 203.81 to 165.19 MPa.

Perspective

The evaluation is retrospective with a fixed candidate pool, and strength values for selected conditions come from an existing database rather than newly fabricated and tested alloys; the conclusions are limited to those 501 records, 14 composition and processing inputs, and ultimate tensile strength as the target.

The authors note that prospective active-learning campaigns, transfer to other alloy families and laboratory-level savings remain to be established; the current diagnostics assess the framework as a whole, leaving the effect of individual components to targeted ablation studies.