MetaForge: An Intelligent Discovery Platform for High-Entropy Alloys Based on Machine Learning and Genetic Algorithms

MetaForge is an end-to-end computational materials science platform designed to address the core challenge of high-entropy alloy (HEA) R&D—its enormous composition space. By integrating Materials Project data, physics-informed machine learning, genetic algorithm-based inverse design, and CHGNet atomic potential relaxation, it enables efficient discovery of optimal HEAs for aerospace, corrosion resistance, refractory, and lightweight structural applications. The project also provides a React+Flask-based interactive web app, allowing researchers to explore HEA design space without coding.

HEAs are composed of 5+ elements in near-equal proportions, offering superior properties like high strength, corrosion resistance, and high-temperature stability. However, their composition space is extremely large—choosing 5 from 20 common transition metals yields over 15,000 theoretical combinations. Traditional trial-and-error methods are ineffective here, necessitating smarter, data-driven approaches.

MetaForge's architecture consists of 6 key modules:

1. Data Layer: Integrates Materials Project API to obtain key physical parameters (atomic radius, density, VEC) for model training.
2. Combination Engine: Generates theoretical alloy formulas and filters via physical rules (lattice strain δ <6.6% for single-phase solid solution; VEC to predict crystal structure).
3. Feature Engineering: Uses Matminer to compute 132 Magpie descriptors covering atomic, thermodynamic, and electronic features.
4. ML Models: Two random forest regression models predict density (RMSE=0.073 g/cm³) and shear strength (RMSE=0.539 GPa) with good interpretability and robustness.
5. Genetic Algorithm: Inverse design optimizes specific strength (strength/density) over 20 generations.
6. Structure Relaxation: Uses CHGNet (graph neural network) to relax 54-atom supercells and output CIF files for further analysis.

MetaForge includes a user-friendly web app with:

• Interactive element proportion sliders
• Real-time ML performance predictions
• Composition pie charts
• CIF file export

A typical optimization result: Alloy Formula: W₀.₁₀Mo₀.₄₀Ta₀.₀₅Nb₀.₀₅V₀.₄₀ Predicted Performance: Density=9.68 g/cm³, Shear Strength=90.71 GPa, Specific Strength=9.37 GPa·cm³/g (leading in refractory HEAs, BCC structure for high-temperature applications).

The app is deployed on Render (free tier; first access may take ~1 minute to wake up).

HEAs' superior properties stem from 4 core effects:

1. High Entropy Effect: Stabilizes single-phase solid solutions via high mixing entropy.
2. Lattice Distortion: Enhances strength via atomic size mismatch.
3. Slow Diffusion: Improves high-temperature creep resistance.
4. Cocktail Effect: Synergistic performance beyond individual elements.

HEAs have applications in aerospace (high-temperature alloys), energy (nuclear reactor parts), chemicals (corrosion-resistant materials), and biomedicine (implants).

MetaForge represents the shift to materials informatics: traditional linear R&D (10-20 years) is replaced by data-driven, AI-powered workflows (months) thanks to big data (Materials Project), advanced algorithms, cloud computing, and open science.

Current limitations:

• Model Precision: Random forest may lag behind deep learning in handling complex nonlinear relationships.
• Data Coverage: Limited to Materials Project data (lacks emerging combinations like high-entropy ceramics).
• Experimental Validation: No integration with high-throughput experimental platforms.
• Single Objective: Optimizes only specific strength (real-world design needs multi-objective tradeoffs).

Future directions:

• Integrate advanced graph neural networks (e.g., Matformer, ALIGNN) for better precision.
• Expand data coverage to emerging material types.
• Link to experimental platforms for validation.
• Adopt multi-objective optimization (e.g., NSGA-II) for balanced performance.

MetaForge revolutionizes HEA discovery by combining AI and materials science into an automated pipeline from theory to valid crystal structures. It provides a practical tool for researchers to accelerate HEA R&D and demonstrates the power of data-driven, open collaboration in materials science. As AI and computing power advance, such platforms will drive a new era of human-AI collaborative material discovery.