Sample-efficient prediction of high-entropy alloy strength through outcome-aware robust sequential learning

Ahmed Shoyeb Raihan, Md Habibor Rahman, Ashish Nimbarte, Imtiaz AhmedarXiv 2026 · arXiv

TL;DR

ROSE-HEA is a new sequential learning framework that significantly reduces the amount of experimental data needed to predict the strength of high-entropy alloys (HEAs). By using a novel "outcome-aware" acquisition policy that minimizes worst-case predictive risk across the entire design space, it achieves higher accuracy with 32% fewer samples than state-of-the-art conventional machine learning models.

What Problem Does This Paper Address?

High-entropy alloys (HEAs) offer a vast design space for structural materials, but exploring this space is a logistical nightmare. Determining the ultimate tensile strength (UTS) of a new alloy requires melting, casting, thermomechanical processing, and destructive testing—a process that is both expensive and time-consuming.

While machine learning (ML) can help predict these properties, standard models require large datasets that simply don't exist for many HEA systems. Existing "active learning" techniques, like Bayesian Optimization (BO), often focus on finding the best material (optimization) rather than building a globally accurate predictive model. The authors address this gap by creating a framework specifically designed for global predictive accuracy under a very tight experimental budget.

Methods

The authors introduce ROSE-HEA (Robust Outcome-aware Sequential Experimentation for High-Entropy Alloys). The framework operates in a closed-loop cycle, illustrated in Figure 1:

The ROSE-HEA closed-loop sequential learning process.

Key technical innovations include:

  • Adaptive Predictive Representation: Instead of relying on a single model, it uses a mixture of three Gaussian Process (GP) hypotheses with different kernels (Matérn 5/2, 3/2, and RBF). This allows the system to remain flexible regarding how "smooth" the underlying physical relationships are.
  • Outcome-Aware Acquisition: Most acquisition functions look at the current state of the model. ROSE-HEA "looks ahead" by simulating possible UTS outcomes for candidate experiments and estimating how those outcomes would update the model's global knowledge.
  • Robust Risk Reduction: The framework uses "distributionally robust" selection. It identifies regions of the design space where the model is most uncertain (the "worst-case" risk) and prioritizes experiments that resolve that specific uncertainty, preventing the model from becoming overly biased toward well-sampled regions.

Datasets

The study performed a retrospective evaluation using a database of 501 HEA records originally reported by Wang et al. (2023).

  • Inputs: 14 variables, including the atomic percentages of 10 elements (C, Al, V, Cr, Mn, Fe, Co, Ni, Cu, Mo) and 4 processing conditions (homogenization temperature, cold-rolling reduction, annealing temperature, and duration).
  • Target: Ultimate Tensile Strength (UTS), which ranged from 320 to 1867 MPa in the dataset.

Key Insights and Ideas

  • Global over Local: Unlike standard BO methods (like Expected Improvement) which "greedily" hunt for the highest strength values, ROSE-HEA is greedy for information. It seeks to improve the model everywhere, not just at the peaks.
  • Transferable Value: A major "aha" moment in the paper is that the samples selected by ROSE-HEA are inherently more informative. When the authors trained other models (like Random Forests or Neural Networks) on the data ROSE-HEA picked, those models also performed better than if they were trained on randomly selected data.
  • Better Coverage: ROSE-HEA naturally pushes experiments toward the "periphery" and sparse areas of the design space. Random sampling tends to cluster in high-density regions of the existing dataset, which wastes resources on redundant information.

Key Results

ROSE-HEA consistently outperformed both standard Bayesian Optimization and conventional ML regressors across 30 matched runs.

  • Accuracy: With a budget of 125 observations, ROSE-HEA achieved a mean Root-Mean-Square Error (RMSE) of 152.84 MPa. This was a 6.16% improvement over standard BO (EI-BO) and a massive 17.86% improvement over the strongest conventional regressor, XGBoost.
  • Sample Efficiency: ROSE-HEA reached a higher level of accuracy with 125 samples than XGBoost could achieve with 185 samples. This represents a 32.4% reduction in the required training budget.
  • Consistency: As shown in the boxplots below, ROSE-HEA (in red) shows a tighter distribution and lower mean error compared to other sequential learning and conventional ML methods.
Figure 2: Prediction at 125 training observations across 30 runs: (a) RMSE and (b) R2. Boxes show interquartile ranges, lines indicate medians, diamonds denote means, and whiskers follow the 1.5-IQR rule. The dotted separator distinguishes SL and conventional ML methods.

The learning progression (Figure 3) highlights that ROSE-HEA’s advantage becomes even more pronounced as more data is acquired, showing a steeper improvement curve than its competitors.

Sequential learning progression from 25 to 125 total observations.

Limitations and Future Work

  • Retrospective Nature: The study used an existing dataset. While this is standard for validating new algorithms, real-world "in-the-loop" experimental campaigns are needed to confirm these savings in a lab setting.
  • Fixed Candidate Pool: The current framework selects from a pre-defined list of candidates. Future iterations could integrate with generative models to propose entirely new alloy compositions.
  • Cost-Agnostic: The model treats all experiments as having equal cost. In reality, some alloy compositions or processing steps are much more expensive or difficult than others. Incorporating "cost-aware" logic would be a valuable next step.

Summary updated

Abstract

Accurate prediction of high-entropy alloy (HEA) strength can guide composition and processing decisions; however, obtaining tensile-test data for HEAs can be costly and experimentally demanding. This study proposes Robust Outcome-aware Sequential Experimentation for High-Entropy Alloys (ROSE-HEA), a sample-efficient framework for predicting ultimate tensile strength (UTS) from alloy composition and processing conditions. Its acquisition policy anticipates how possible UTS outcomes update predictions and the support for complementary Gaussian process (GP) descriptions. It then reassesses the worst-case expected predictive risk remaining across a reference domain to select informative candidates. A retrospective evaluation on 501 HEA records compares ROSE-HEA with expected-improvement (EI) and upper-confidence-bound (UCB) Bayesian optimization (BO), random GP sampling, and eight conventional regressors across 30 matched runs with shared initial observations and test conditions. Starting from 25 initial observations and acquiring 100 additional responses, ROSE-HEA achieves the lowest mean root-mean-square error (RMSE), at 152.84 MPa, and the highest mean R^2, at 0.7443. Mean RMSE is reduced by 6.16% and 6.37% relative to EI and UCB, respectively. The selected observations improve prediction across six regressor families and provide broader design-space coverage. ROSE-HEA also surpasses the strongest conventional regressor's mean predictive performance using 125 rather than 185 observations, corresponding to 32.4% fewer training observations. These findings demonstrate the potential of outcome-aware robust sequential learning to improve UTS prediction in HEAs under constrained training budgets.
BibTeX
@misc{raihan2026sample,
  author = {Ahmed Shoyeb Raihan and Md Habibor Rahman and Ashish Nimbarte and Imtiaz Ahmed},
  title = {{Sample-efficient prediction of high-entropy alloy strength through outcome-aware robust sequential learning}},
  year = {2026},
  eprint = {2611.18462},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18462},
}
Figure