Accelerating Alloy Discovery with Bayesian Optimization

By Abdul Rahman · June 15, 2026 · Computational Materials Science

How we combine physical metallurgy constraints with active learning to discover high-performance structural steels and Zn-based alloys in minimal experimental iterations.

The Challenge of Modern Alloy Design

Traditional metallurgical exploration relies heavily on Edisonian trail-and-error loops—fabricating dozens of alloy samples, heat-treating them under varied schedules, and performing tedious mechanical characterization. This physical limit restricts materials scientists to a fraction of the vast multi-element compositional space.

In physical metallurgy, finding the optimal balance of tensile strength, elongation, and environmental degradation resistance is inherently a multi-objective optimization problem.

Enter Bayesian Optimization

To accelerate this bottleneck, we integrated Bayesian Optimization coupled with gaussian process regressors to intelligently guide our physical experiments.

Rather than randomly exploring compositions, our active learning loop works as follows:

1. Prior Knowledge Initialization: We seed our model with historical data from literature and thermodynamic databases (CALPHAD).
2. Surrogate Modeling: A Gaussian Process model predicts performance across thousands of prospective candidate configurations, providing both a predicted mean (expected strength) and uncertainty (variance).
3. Acquisition Function: Using Expected Improvement (EI), the system selects the next specific composition to physically formulate at our PCSIR facility. This balances exploration (high uncertainty regions) with exploitation (high predicted performance).

```python

Example of Bayesian Active Learning Step

from sklearn.gaussian_process import GaussianProcessRegressor from sklearn.gaussian_process.kernels import Matern

Multi-objective compositional optimization

kernel = Matern(nu=2.5) gp = GaussianProcessRegressor(kernel=kernel, alpha=1e-5, n_restarts_optimizer=10) gp.fit(X_train, y_train) # X: Alloy composition %, y: Yield Strength ```

Impact & Results

Using this informatics-driven approach on Zn-based alloy systems, we reduced the number of experimental castings required to achieve a targeted tensile threshold by over 60%. Furthermore, physical metallurgy constraints (such as preserving phase balance like ferrite/austenite ratios in weld cladding) were incorporated directly as penalty constraints in the acquisition function.

The convergence of computational Materials Informatics and classical laboratory expertise is no longer a future concept—it is the modern standard for accelerating industrial discovery.