Fine-tuning of universal machine-learning interatomic potentials for 2D high-entropy alloys

This study demonstrates that fine-tuning universal machine-learning interatomic potentials on systematically generated structures enables near-DFT accuracy in predicting mixing energies for 2D high-entropy alloys, overcoming the computational limitations of direct DFT calculations for complex systems like experimentally synthesized (Mo,Ta,Nb,W,V)S$_2$.

The Gist

Imagine you are a chef trying to create the ultimate "Super-Salad." You want to mix five different types of vegetables (let's say Mo, Ta, Nb, W, and V) into a single bowl to get a flavor that is stronger, more stable, and more nutritious than any single vegetable alone. This is essentially what scientists are doing with High-Entropy Alloys (HEAs): mixing multiple metals together to create materials with amazing new properties, like super-catalysts for cleaning our air or generating clean energy.

However, there's a huge problem: The kitchen is too small.

The Problem: The "Super-Computer" Kitchen

To figure out exactly how these metals behave when mixed, scientists usually use a powerful tool called Density Functional Theory (DFT). Think of DFT as a hyper-accurate, slow-motion food critic who tastes every single atom in your salad to tell you if it's good.

The problem is that for a complex salad with five ingredients, the number of possible ways to arrange them is astronomical. Running the "food critic" (DFT) on every possible arrangement would take a supercomputer thousands of years. It's simply too expensive and slow.

The Shortcut: The "Universal Recipe Book"

To speed things up, scientists started using Machine Learning Interatomic Potentials (MLIPs). Imagine these as a "Universal Recipe Book" trained on millions of different dishes. Instead of tasting every single atom, the book predicts the flavor based on patterns it has seen before.

Recently, scientists created Universal MLIPs (like MACE, MatterSim, and CHGNet). These are like a massive, general-purpose cookbook that claims to know how to cook anything—from steak to sushi to your five-vegetable salad.

But here's the catch: When the scientists tried to use these "Universal Cookbooks" to predict the flavor of their specific 5-vegetable salad, the results were terrible. The book guessed the taste was completely wrong. It was like a cookbook trained on Italian food trying to predict the taste of a spicy Thai curry; it just didn't have the specific details needed.

The Solution: "Fine-Tuning" the Recipe

This is where the paper comes in. The authors asked: "Can we take this Universal Cookbook and teach it specifically how to make our 5-vegetable salad?"

They developed a strategy called Fine-Tuning.

1. The Old Way (Random Sampling): You could try to teach the book by showing it random, messy piles of vegetables. This helps it get the "average" taste, but it might miss the rare, perfect combinations.
2. The New Way (Enumeration): Instead of guessing randomly, the authors systematically listed every possible way to arrange the vegetables in a small bowl (like a grid). They showed the book every single permutation: Mo-Ta-Nb, Ta-Mo-V, etc.

The Analogy:

• Random Training is like showing a student a few random photos of dogs. They might learn what a dog looks like on average, but they might fail if shown a very specific breed they've never seen.
• Enumeration Training is like showing the student a complete encyclopedia of every dog breed, from Chihuahuas to Great Danes. They learn the rules of "dog-ness" perfectly.

The Results: A Master Chef

The team found that by fine-tuning the Universal Cookbook with these systematically listed (enumerated) structures, they created a model that was:

• As accurate as the slow, expensive food critic (DFT).
• Fast enough to run millions of simulations in seconds.

They used this new "Master Chef" model to simulate what happens to their 5-metal salad when you heat it up. They discovered that at around 400°C, the salad starts to fall apart. Specifically, one ingredient (Vanadium) really doesn't want to mix with the others and tries to leave the bowl, while the others stick together.

This matches real-world experiments perfectly, proving their method works.

Why This Matters

This paper is a game-changer because it gives scientists a fast, cheap, and accurate way to design new materials.

• Before: Designing a new alloy was like trying to find a needle in a haystack by looking at one straw at a time with a magnifying glass.
• Now: With these fine-tuned models, they have a metal detector that can scan the whole haystack in minutes.

In summary: The authors took a "one-size-fits-all" AI tool, taught it the specific rules of a complex metal salad using a systematic list of all possibilities, and turned it into a super-powerful tool that can predict how these materials behave without needing a supercomputer for every single calculation. This opens the door to designing better batteries, catalysts, and materials for the future much faster than ever before.

Technical Summary

1. Problem Statement

High-entropy alloys (HEAs) and their two-dimensional counterparts (2D-HEAs) offer tunable properties and significant catalytic potential. However, their chemical complexity (multiple principal elements in near-equimolar ratios) creates two major computational bottlenecks:

• DFT Limitations: Direct Density Functional Theory (DFT) calculations are computationally prohibitive due to the need for large supercells to capture chemical disorder and the vast configurational/compositional space requiring ensemble averaging.
• MLIP Limitations: While Machine Learning Interatomic Potentials (MLIPs) offer a solution, Universal MLIPs (uMLIPs) trained on general databases (e.g., MPtrj) often yield unsatisfactory predictions for mixing energies in complex HEA systems without further adaptation.

The core challenge addressed is how to effectively adapt (fine-tune) existing universal models to accurately predict the energetics of 2D transition metal dichalcogenide (TMDC) HEAs, specifically the experimentally synthesized (Mo,Ta,Nb,W,V)S₂ system, while maintaining computational efficiency for large-scale simulations like Monte Carlo (MC).

2. Methodology

A. Computational Framework

• DFT Reference: Calculations were performed using VASP with the PBE functional and PAW pseudopotentials. Parameters were strictly aligned with the Materials Project to ensure consistency with the foundation models.
• Universal Models Tested: Five leading uMLIPs were benchmarked: MACE-small, MACE-small-0b2, MatterSim-1m, MatterSim-5m, and CHGNet.
• Fine-Tuning Strategy: The authors compared two distinct approaches for generating training sets:
  1. Random Structures: Randomly selected configurations (common in previous HEA studies).
  2. Enumerated Structures: Systematically generated structures using the Interaction Cluster Expansion Toolkit (ICET) to exhaustively sample symmetry groups and stoichiometries.
• Training Protocol: Models were fine-tuned using 80% training, 15% testing, and 5% validation splits. The loss function weighted energy, forces, and stress at a 100:10:1 ratio.

B. Benchmarking Databases
Three distinct databases were constructed to evaluate model performance:

1. Database 1: Binary alloys at five concentrations (150 structures).
2. Database 2: Random concentrations/configurations for 2-, 3-, 4-, and 5-component alloys (80 structures).
3. Database 3: Enumerated equimolar structures for $N$ components ($N \le 5$).

C. Application
The best-performing fine-tuned model was applied to:

• Phase Decomposition: Calculating Gibbs free energy to predict decomposition products.
• Monte Carlo Simulations: Using the MCHAMMER package to study temperature-dependent phase separation and cluster vectors (CV).
• Synthesizability: Evaluating the Disordered Enthalpy–Entropy Descriptor (DEED).

3. Key Contributions

1. Benchmarking uMLIPs on 2D HEAs: The study demonstrates that while uMLIPs predict total energies well, they fail significantly in predicting mixing energies for HEAs without fine-tuning, often showing excessive dispersion or deviation from DFT.
2. Systematic Comparison of Training Strategies: The paper provides a rigorous comparison between random and enumerated training structures. It establishes that while random structures may yield lower Mean Absolute Errors (MAE) on average, enumerated structures provide superior transferability and safety by covering the configurational phase space more comprehensively (avoiding "blind spots" in ordered structures).
3. Optimization of Training Sets: The authors identified that a training set consisting of all possible element combinations (even with fewer components) is sufficient to model complex HEAs. Specifically, they found that a model trained on a complete set of binary and ternary enumerated structures (Model 4) achieved accuracy comparable to a model trained on the full 5-component dataset (Model 16), but with significantly fewer structures.
4. Fine-tuning vs. Training from Scratch: The study shows that fine-tuning is superior to training from scratch for small-to-medium datasets. Models trained from scratch on limited data failed to relax structures stably, whereas fine-tuned models remained robust.

4. Key Results

• Accuracy Improvement: Fine-tuning reduced the MAE for mixing energy from 25 meV/unit (foundation MACE-0b2) to **1.5–2.0 meV/unit**, achieving near-DFT accuracy.
• Model Performance:
  • MACE-based models (specifically MACE-small-0b2) were selected as the best foundation models due to a balance of accuracy and stability in molecular dynamics.
  • Enumerated models outperformed random models in transferability. A model trained on enumerated structures (Model 16) achieved the lowest errors across all databases.
  • Data Efficiency: A model trained on a subset of enumerated structures (Model 4, ~31% of the size of Model 16) achieved convergence-level accuracy, proving that structural diversity (covering all element pairs) is more critical than sheer volume.
• Phase Decomposition of (Mo,Ta,Nb,W,V)S₂:
  • Decomposition Temperature: The alloy is predicted to decompose around 400 K.
  • Products: The system separates into (Mo,W)S₂, (Ta,Nb)S₂, and VS₂.
  • V Behavior: Vanadium (V) strongly disfavors mixing with other elements, leading to early separation (even at higher temperatures in MC simulations), consistent with experimental observations where V has the lowest atomic fraction in synthesized samples.
• DEED Value: The calculated Disordered Enthalpy–Entropy Descriptor (DEED) was 247 (eV/atom)⁻¹, well above the threshold for synthesizability (~35 for borides), confirming the experimental success of synthesizing this 2D HEA.

5. Significance

• Methodological Advancement: This work establishes enumerated structure fine-tuning as a robust, efficient strategy for adapting universal MLIPs to complex, multi-component systems. It resolves the trade-off between data volume and structural diversity.
• Scalability: The fine-tuned models enable Monte Carlo simulations on scales (large supercells, long timescales) that are inaccessible to DFT, allowing for the study of temperature-dependent phase evolution and short-range order.
• Material Discovery: The successful prediction of phase stability and decomposition behavior for (Mo,Ta,Nb,W,V)S₂ validates the approach for designing new 2D HEAs for catalysis (HER, CO₂ reduction) and tribological applications.
• Generalizability: The findings suggest that similar fine-tuning strategies can be applied to other complex alloys beyond HEAs, offering a pathway to accelerate materials discovery where chemical complexity hinders traditional simulation methods.

In conclusion, the paper demonstrates that while universal MLIPs are not "plug-and-play" for HEAs, a targeted fine-tuning strategy using systematically enumerated training data can produce highly accurate, stable potentials capable of guiding the design and understanding of complex 2D high-entropy materials.