🔬 materials science

Exploring the extremes: atomic basis for multi-elemental materials science under complex thermodynamic conditions

This paper introduces a chemistry-agnostic, information-entropy-maximization protocol for generating training data that overcomes the limitations of current machine-learning interatomic potentials in complex, multi-elemental systems, thereby enabling robust and unbiased simulation of materials under extreme thermodynamic conditions.

Original authors: Anton Bochkarev, Yury Lysogorskiy, Aparna Subramanyam, Ralf Drautz, Danny Perez

Published 2026-03-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Anton Bochkarev, Yury Lysogorskiy, Aparna Subramanyam, Ralf Drautz, Danny Perez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: From "Recipe Books" to "The Whole Kitchen"

Imagine you are a chef. For the last 100 years, you've only cooked using a tiny, specific set of ingredients: flour, sugar, eggs, and butter. You've mastered these recipes perfectly. You know exactly how a cake rises or how bread browns. This is how modern materials science has worked. We mostly study materials made of just a few elements (like steel, which is mostly iron and carbon) because they are easy to understand and safe to predict.

But the world is changing. We need to recycle everything (the "circular economy"), build machines that survive nuclear explosions, or design materials for other planets. These situations require mixing everything in the periodic table—maybe 20, 30, or even 90 different elements at once.

The problem? Our old "recipe books" (databases) are useless here. They only have recipes for cakes and cookies. If you try to cook a "Mendeleev stew" (a mix of almost all elements) using those old books, your computer simulation will crash or give you nonsense results.

This paper introduces a new way to build a recipe book that covers the entire kitchen.

The Problem: The "Safe Zone" Trap

Most computer models for materials are trained on data from "safe zones." Think of this like a driving school that only teaches you how to drive on a sunny day on a straight, empty highway.

The Training Data: Scientists usually feed computers data about materials that are calm, stable, and at room temperature.
The Failure: When you ask the computer, "What happens if I drive this car off a cliff at 200 mph in a hurricane?" (which is what happens in a fusion reactor or a nuclear explosion), the computer panics. It has never seen that data before. It guesses wildly and gets it wrong.

The authors found that even the biggest, most famous databases (like OMAT, which has millions of entries) are still missing huge chunks of the "driving map." They are missing the cliffs, the hurricanes, and the weird chemical combinations.

The Solution: The "Maximum Entropy" Chef

To fix this, the authors invented a new method called SMAX (Statistically Maximizing Entropy).

The Analogy: The Blindfolded Chef
Imagine you want to learn every possible way to arrange ingredients in a bowl.

Old Way: You ask a human chef, "What are the best arrangements?" The chef only gives you the ones that taste good (low energy, stable). You miss the weird, messy, or broken arrangements that might happen in an accident.
The SMAX Way: You blindfold the chef. You tell them, "Don't worry about taste. Just make sure you try every single possible arrangement of atoms, from the most stable to the most chaotic, without favoring any specific one."

By using a mathematical trick called Information Entropy, the computer generates structures that are chemically diverse but not biased toward being "stable." It forces the system to explore the weird, the wild, and the extreme.

The Result: They created a database with nearly 1.7 million structures containing over 23 million atoms. It covers almost the entire periodic table (except the very radioactive ones). It's like having a map of the entire universe of atoms, not just the "nice neighborhoods."

The Test: Putting the New Map to Work

They trained a new AI model (called GRACE) on this new "SMAX" map and put it to the test against the old models. Here is how it performed in three real-world scenarios:

1. The "Crush Test" (Tin Deformation)

The Scenario: Imagine taking a block of tin and crushing it until it changes shape completely.
The Old Model: It got confused. It thought the tin would just break or behave like it was in a calm lab.
The SMAX Model: It predicted the exact new shape the tin would take under extreme pressure. It knew how the atoms would rearrange because it had "seen" similar crushing scenarios during its training.

2. The "Radiation Damage" Test (Tungsten Alloys)

The Scenario: In a nuclear fusion reactor, materials get hit by high-speed particles, creating tiny holes (defects) and chaos in the atomic structure.
The Old Model: It failed to predict how these holes would move or how the material would weaken. It was like a doctor who only studied healthy people and couldn't diagnose a broken bone.
The SMAX Model: It accurately predicted how the defects would behave, even in complex alloys made of 5 different metals. It handled the "messy" atomic neighborhoods perfectly.

3. The "Lava" Simulation (Discovery by Simulation)

The Scenario: This was the ultimate test. They took the 9 most common elements in the Earth's crust (Oxygen, Silicon, Iron, etc.) and mixed them up randomly. Then, they simulated cooling them down from the heat of a volcano to room temperature.
The Result: The AI didn't just predict a known material. It discovered new structures on its own!
- It spontaneously formed metallic clusters of Iron and Silicon inside an oxide matrix.
- It figured out how different elements would separate (segregate) based on their chemistry.
- It did this without the scientists telling it what to look for. The AI just "saw" the patterns because its training data was so broad.

Why This Matters: The "Discovery by Simulation" Era

The biggest takeaway is a shift in how we do science.

Old Way: "I have a hypothesis. I will build a model to test if my specific idea works." (Like checking if a specific key opens a specific lock).
New Way: "I will let the AI explore the entire universe of possibilities, and it will tell me what structures naturally emerge." (Like letting a river flow and seeing where it carves a new path).

This approach allows us to design materials for:

Fusion Energy: Materials that can survive the sun's heat.
Recycling: Understanding how to separate complex electronic waste (which contains 60+ elements) without knowing the exact recipe beforehand.
Space Exploration: Materials that can handle the extremes of other planets.

The Bottom Line

The authors realized that the problem wasn't that our AI models were too "dumb." The problem was that we were feeding them a diet of only "healthy, stable food." By feeding them a "maximum entropy" diet—covering every weird, chaotic, and extreme possibility—we gave them the intuition to handle the real, messy world.

They didn't just build a better map; they built a compass that works even when the terrain is unknown.

1. Problem Statement

Modern materials science has historically relied on systems with limited chemical complexity (typically 3–4 major elements) and near-equilibrium conditions. However, emerging challenges—such as the circular economy (recycling complex waste streams), nuclear fusion (extreme radiation environments), and planetary science—require the understanding of "Mendeleev materials": systems containing a vast, non-uniform mixture of many elements (potentially spanning the entire periodic table) under extreme thermodynamic conditions (high pressure, temperature, irradiation).

Current Universal Machine Learning Interatomic Potentials (MLIPs) fail in these regimes because:

Data Bias: Traditional training datasets (e.g., OMAT, Materials Project) are heavily biased toward low-energy, near-equilibrium crystal structures.
Systematic Gaps: They lack coverage of far-from-equilibrium configurations, high-energy transition states, and complex multi-elemental disorder.
Extrapolation Failure: When tested on extreme conditions or complex mixtures, these models suffer catastrophic errors because they attempt to extrapolate beyond their training distribution.

2. Methodology

The authors propose a paradigm shift from "human-curated, low-energy" data generation to a chemistry-agnostic, information-entropy maximization protocol.

A. The SMAX Dataset Generation

Core Concept: Instead of targeting specific materials, the authors maximize the information entropy of the dataset in a chemically meaningful feature space. This decouples structural sampling from thermodynamic bias.
Algorithm: A greedy optimization scheme iteratively generates atomic configurations. At each step, a candidate structure is optimized to maximize the determinant of the feature covariance matrix ( $\log \det(\Sigma)$ ), ensuring the new configuration adds maximum diversity to the existing dataset.
Constraints: The process includes physical constraints (minimum interatomic distances, minimum density) to prevent unphysical overlaps but remains agnostic to specific chemical identities or energy minima.
Scope: The resulting SMAX (Statistically Maximized Entropy) dataset contains 1.69 million structures with 23 million atoms, covering nearly the entire periodic table (excluding actinides) and spanning a vast range of energies, forces, and stresses.
Enrichment: The core SMAX set is enriched with low-energy structures from existing databases (Materials Project, QM9) to ensure coverage of ground states, but the dominant feature is the systematic sampling of the "unexplored" high-entropy space.

B. Model Architecture: GRACE

The authors train Graph Atomic Cluster Expansion (GRACE) models on the SMAX dataset.

GRACE: A state-of-the-art MLIP framework that extends the Atomic Cluster Expansion (ACE) to graph-based functions, allowing for the systematic inclusion of semilocal and nonlocal interactions.
Training Strategy: Three models were compared:
1. SMAX Model: Trained exclusively on the entropy-maximized dataset.
2. OMAT Model: Trained on the massive, conventional OMAT dataset (baseline).
3. SMAX–OMAT Model: Trained on a combination of both.

3. Key Contributions

SMAX Dataset: The creation of the first dataset explicitly designed to systematically cover the chemical and configurational space of the periodic table via entropy maximization, rather than thermodynamic relevance.
Chemistry-Agnostic Sampling: Demonstrating that decoupling data generation from specific material energetics creates a robust "physical prior" for atomic interactions.
Discovery by Simulation: Establishing a workflow where unbiased simulations can autonomously resolve emergent structures in multi-elemental mixtures without a priori chemical assumptions.

4. Key Results

A. Benchmarking and Transferability

Failure of OMAT-trained Models: When evaluated on the SMAX test set, the OMAT-trained GRACE model showed catastrophic failure (Force MAE $\approx$ 3 eV/Å, Stress MAE > 9,000 GPa), despite performing well on its own hold-out set. This proves that dataset size alone does not guarantee universality.
Robustness of SMAX Models: The SMAX-trained model achieved balanced, low error rates across both the SMAX and OMAT test sets (Force MAE $\approx$ 0.14 eV/Å), demonstrating superior generalization and the elimination of "blind spots" in the potential energy surface.

B. Application Benchmarks

Tin (Sn) Large Deformation:
- The SMAX model accurately reproduced complex energy landscapes for phase transitions (diamond-cubic $\to$ $\beta$ -Sn) under extreme uniaxial compression.
- OMAT models showed systematic deviations in saddle regions and high-strain basins.
Defects in Tungsten-Based Materials (Fusion Context):
- W/Be Interdiffusion: The SMAX model correctly predicted formation energies for intermetallic compounds and defects under full relaxation, whereas OMAT models failed to stabilize mechanically unstable configurations.
- High-Entropy Alloys (HEAs): In W-Ta-Cr-V alloys, the SMAX model provided the lowest error for vacancy formation energies in chemically disordered environments, crucial for predicting radiation damage.
Grain Boundary Segregation:
- Successfully simulated complex segregation patterns in refractory HEAs (W-Nb-Mo-Ta-V) with light impurities (H, B, C, N), capturing the competition between metallic demixing and light-element partitioning.
Catalytic Reaction Barriers:
- On the OC20NEB benchmark (out-of-distribution reaction barriers), the SMAX-based models significantly outperformed OMAT baselines, despite never being explicitly trained on transition states.
Autonomous Structure Formation:
- "Lava" Simulation: Simulating a mixture of the 9 most abundant crustal elements resulted in the spontaneous nucleation of Fe-Si metallic clusters within an oxide matrix.
- "Mendeleev" Simulation: A mixture of 94 elements spontaneously segregated into distinct phases (refractory carbides, actinide-iron eutectics, stable fluorides, and radioactive salts) based purely on physical interactions, validating the "discovery by simulation" paradigm.

5. Significance and Conclusion

This work fundamentally changes the approach to materials discovery:

From Hypothesis-Driven to Data-Driven: It moves away from testing specific hypotheses about known materials toward autonomously exploring the vast, uncharted chemical space.
Solving the "Universal" Potential Problem: It demonstrates that the bottleneck for universal MLIPs is not model architecture, but dataset bias. By maximizing information entropy, the authors provide a rigorous foundation for simulating materials under extreme conditions (fusion, planetary interiors) and complex industrial scenarios (waste recycling).
Scalability: The methodology is scalable and principled, enabling the simulation of "Mendeleev materials" that were previously computationally intractable due to the combinatorial explosion of possible configurations.

In summary, the paper establishes that systematic, entropy-maximized sampling is the key to unlocking reliable, universal atomic simulations for the next generation of complex, multi-elemental materials science.