Inverse Design of Aluminium Alloys Using Genetic Algorithm: A Class-Based Workflow

Ninad Bhat, Amanda S. Barnard, N. BirbilisMetals 2024 · DOI

TL;DR

This paper introduces a novel "class-based" approach to inverse alloy design for aluminum alloys, using machine learning and genetic algorithms. By training separate models for distinct alloy classes, they achieve better predictions of high-strength and high-elongation alloys compared to traditional methods that treat all alloys as one group. This work offers a more refined pathway to discovering new, high-performance aluminum alloys.

What Problem Does This Paper Address?

Designing new aluminum alloys is a complex challenge due to the vast number of alloying elements, their varying concentrations, and different processing conditions. A key hurdle is the inherent trade-off between strength and ductility – making an alloy stronger often makes it more brittle. Traditional "inverse design" methods, which aim to find alloy compositions that meet target properties, typically use a machine learning model to predict properties and a genetic algorithm to search for optimal compositions. However, these methods often treat all alloys as a single, homogenous group, which can be inefficient and lead to suboptimal results. This paper addresses the limitation of class-agnostic approaches by proposing a class-based workflow to improve the accuracy and efficiency of inverse alloy design.

Methods

The core of this research lies in a "class-based" workflow for inverse alloy design. Here's a breakdown of their methodology:

  • Data-Driven Partitioning: The initial step involves clustering the dataset of aluminum alloys into distinct classes based on their inherent similarities. This is achieved using unsupervised machine learning techniques, specifically iterative label spreading, to identify eight separable classes.
  • Class-Specific Forward Prediction: For each identified class, individual random forest regressors are trained. Random forests are ensemble learning methods known for their accuracy in predicting complex relationships. Crucially, before training, recursive feature elimination is used to identify and retain only the most important features (alloying elements and processing conditions) for each class, reducing noise and improving model efficiency.
  • Class-Based Inverse Design: Once the class-specific regressors are trained, a multi-objective genetic algorithm (specifically NSGA-II) is employed. Instead of searching across the entire alloy space, the genetic algorithm is applied independently to each class, using its corresponding class-specific regressor to evaluate potential alloy compositions. The objectives are to maximize both tensile strength and elongation.
  • Comparison: The Pareto fronts (sets of optimal trade-offs) generated by the class-based approach are compared against a Pareto front generated by a single, class-agnostic random forest model trained on the entire dataset. This comparison aims to demonstrate the superiority of the class-based method.

Datasets

The study utilizes a publicly accessible dataset of aluminum alloys curated by the authors. This dataset includes:

  • Alloy Types: Cast and wrought alloys, including age-hardened and strain-hardened types.
  • Mechanical Properties (Targets): Tensile strength (44.81–820 MPa), yield strength (151.26–790 MPa), and elongation (0.5–50%).
  • Features: Concentrations of 25 different alloying elements and ten distinct processing conditions.
  • Instances: The dataset contains 1154 instances, with 933 alloys having complete data for all three mechanical properties.

This dataset is a valuable resource for alloy design research, encompassing a broad range of aluminum alloys and their properties.

Key Insights and Ideas

  • The Power of Specialization: The central idea is that by specializing predictive models to specific classes of alloys, we can achieve more accurate and targeted inverse design. Treating all aluminum alloys as one monolithic group masks important class-specific relationships.
  • Feature Importance Varies by Class: The study highlights that the importance of different alloying elements and processing conditions can vary significantly between alloy classes. This justifies the need for class-specific feature selection and modeling.
  • Reduced Feature Space for Class-Based Models: Recursive feature elimination revealed that class-based regressors often require fewer features to achieve optimal accuracy compared to a class-agnostic model. This suggests that class-specific models are more efficient and less prone to overfitting.
  • Dominant Pareto Fronts: The class-based approach successfully generated Pareto fronts that dominate the Pareto front from the class-agnostic model in specific regions. This means for certain target property ranges, the class-based method found alloys with superior combinations of strength and elongation.
  • Discovery of Novel Compositions: The class-based inverse design identified some potentially novel alloy compositions, particularly in the 7xxx series, with high Zn and Mg concentrations predicted to exceed 700 MPa tensile strength, pushing beyond current literature.

Key Results

  • Improved Prediction Accuracy: While not explicitly quantified as a single metric, the study demonstrates that class-based regressors, after feature selection, generally achieved lower Mean Absolute Error (MAE) and Mean Squared Error (MSE) compared to the class-agnostic regressor trained on the entire dataset. For instance, MAE for Class 4 and Class 8 were significantly lower than the class-agnostic model.
  • Dominant Pareto Fronts: Figure 5 visually shows that Pareto fronts P1, P2, and P6 (derived from class-based optimization) outperform the class-agnostic Pareto front (Pall) in specific regions, offering better trade-offs between tensile strength and elongation.
  • Efficient Feature Selection: Recursive feature elimination identified optimal feature subsets for each class, often requiring fewer features than the full set used by the class-agnostic model. For example, Class 4 required only 2 features for optimal performance.
  • Identification of Promising Alloys: The study predicted specific alloy compositions within each class, some of which align with known alloy types (e.g., AA5xxx for Class 1, 7xxx for Class 2 and 6) and others that represent potentially new compositions for further experimental investigation.

Limitations and Future Work

  • Experimental Validation Needed: The authors acknowledge that the predicted alloys require experimental validation to confirm their properties and suitability for specific applications.
  • Missing Processing Parameters: The current model does not explicitly include critical processing parameters like ageing time and temperature, which are known to significantly influence alloy properties. Incorporating data on these parameters could further enhance the model's applicability.
  • Error Estimation: The forward model provides an estimate of prediction error, but experimental validation is essential to accurately estimate the error in predicted alloy concentrations.
  • Future Work: The authors suggest extending the workflow to include other properties like conductivity or hardness, and augmenting the dataset with calculated parameters such as phase concentrations. Further experimental validation of predicted alloys, especially novel compositions, is a clear next step. The study also points to the need for further research into issues like hot cracking and stress corrosion for high-Zn aluminum alloys.

Summary updated

Abstract

The design of aluminium alloys often encounters a trade-off between strength and ductility, making it challenging to achieve desired properties. Adding to this challenge is the broad range of alloying elements, their varying concentrations, and the different processing conditions (features) available for alloy production. Traditionally, the inverse design of alloys using machine learning involves combining a trained regression model for the prediction of properties with a multi-objective genetic algorithm to search for optimal features. This paper presents an enhancement in this approach by integrating data-driven classes to train class-specific regressors. These models are then used individually with genetic algorithms to search for alloys with high strength and elongation. The results demonstrate that this improved workflow can surpass traditional class-agnostic optimisation in predicting alloys with higher tensile strength and elongation.
BibTeX
@inproceedings{bhat2024inverse,
  author = {Ninad Bhat and Amanda S. Barnard and N. Birbilis},
  title = {{Inverse Design of Aluminium Alloys Using Genetic Algorithm: A Class-Based Workflow}},
  booktitle = {Metals},
  year = {2024},
  doi = {10.3390/met14020239},
}
Figure