Expanding Universal Machine Learning Interatomic Potentials to 97 Elements Towards Nuclear Applications

This paper presents an open-source universal machine learning interatomic potential covering 97 elements, including minor actinides, by integrating a newly constructed heavy element dataset with existing resources to enable advanced materials design for nuclear applications.

The Gist

Imagine you are trying to build a massive, ultra-realistic video game of the universe. To make the physics work, you need a rulebook that tells every single atom how to behave: how they bounce off each other, how they stick together, and how they move when heated up.

For decades, scientists have had a very accurate but incredibly slow rulebook called DFT (Density Functional Theory). It's like calculating the physics of every single raindrop in a storm using a supercomputer. It's perfect, but it takes so long that you can't simulate a whole city, let alone a nuclear reactor.

Then, scientists invented Machine Learning Interatomic Potentials (MLIPs). Think of these as "smart shortcuts." They are like a seasoned chef who has tasted thousands of dishes and can instantly guess the flavor of a new one without cooking it first. These AI models are millions of times faster than the old rulebook but still incredibly accurate.

The Problem: The "Heavy" Missing Pieces
Here's the catch: The best AI chefs (universal MLIPs) had only learned to cook with 89 ingredients (elements). They knew everything about Carbon, Iron, and Gold. But they had never tasted the "heavy" ingredients like Americium, Curium, or Californium.

These heavy elements are crucial for the future of nuclear energy (cleaner fuel, better waste management) and space exploration (powering deep-space probes). But because they are radioactive and dangerous, it's incredibly hard and expensive to test them in a real lab. Scientists were flying blind, unable to simulate how these materials would behave because their AI models didn't know they existed.

The Solution: The "HE26" Cookbook
This paper is about a team of researchers from Osaka University who decided to fix this gap. They created a brand new "cookbook" called HE26.

1. Gathering the Ingredients: They scoured old scientific literature and ran new, complex computer simulations to gather data on 8 rare, heavy elements that were previously missing from the AI's diet.
2. The Big Mix: They didn't just stop at the heavy stuff. They mixed this new "heavy element" data with two massive existing datasets (one for crystals, one for molecules). It's like taking a new, exotic spice and blending it into a giant pot of soup that already contains thousands of other ingredients.
3. Training the AI: They fed this massive, combined dataset (now covering 97 elements) into a new AI model they named MACE-Osaka26.

The Result: A Universal Simulator
The result is a super-AI that can now predict how almost any material in the periodic table will behave.

• The Test Drive: They tested it on things it had never seen before, like complex mixtures of heavy elements used in nuclear fuel. The AI got it right.
• The Heat Test: They used the AI to calculate how heat moves through these nuclear materials. This is vital for designing safe reactors. The AI predicted the heat flow so accurately that it matched real-world experiments, even for materials that are too dangerous to test easily.

Why This Matters (The Analogy)
Imagine you are designing a new type of car engine that runs on a mysterious, super-dense fuel.

• Before this paper: You had a blueprint for engines using gasoline, diesel, and electricity. But for your new fuel, you had to guess how the pistons would move. You were guessing in the dark.
• After this paper: You now have a complete, 3D simulation of how that mysterious fuel interacts with every part of the engine. You can test thousands of designs in a day on a computer, finding the perfect one without ever building a dangerous prototype.

In a Nutshell
This paper gives scientists a "universal translator" for the atomic world. By teaching AI about the heaviest, most radioactive elements, they have unlocked the ability to design the next generation of nuclear energy, space power sources, and advanced materials much faster and safer than ever before. It's a giant leap from "guessing" to "knowing" how the most extreme materials in the universe behave.

Technical Summary

Here is a detailed technical summary of the paper "Expanding Universal Machine Learning Interatomic Potentials to 97 Elements: Towards Nuclear Applications."

1. Problem Statement

• Limitations of Existing Models: While Universal Machine Learning Interatomic Potentials (MLIPs) have advanced significantly, covering up to 89 elements, they lack coverage for heavy elements crucial to nuclear science, specifically minor actinides (e.g., Americium, Curium, Californium) and other transuranium elements.
• Data Scarcity: Existing large-scale first-principles datasets (like MPtrj) have limited or no coverage of these heavy elements due to the extreme difficulty, cost, and radioactivity associated with experimental measurements and the computational complexity of Density Functional Theory (DFT) calculations for strongly correlated $f$-electron systems.
• Application Gap: This gap hinders the computational design of next-generation nuclear materials, such as actinide-based high-entropy ceramics, mixed-oxide fuels, and radioisotope thermoelectric generators, where accurate prediction of thermophysical properties (e.g., lattice thermal conductivity) is essential.

2. Methodology

The authors developed a comprehensive pipeline to construct a new dataset and train a universal model:

• Dataset Construction (HE26):

  • Created a new DFT dataset named HE26 (Heavy Element 2026) containing 65 elements, explicitly adding eight heavy elements rarely found in existing datasets: Am, Cm, Cf, Fr, At, Ra, Po, and Rn.
  • The dataset is divided into three subsets:
    1. BHE (Basic Heavy Element): Elemental solids and binary oxides.
    2. CHE (Complex Heavy Element): Diverse multi-component systems.
    3. CCFO (Compositionally Complex Fluorite Oxides): Binary-to-quinary mixed oxides of actinides, lanthanides, and rare earths, targeting high-entropy ceramics.
  • DFT Protocols: Calculations were performed using VASP with PBE functionals (and DFT+U where necessary). A rigorous magnetic configuration screening process was implemented to handle the complex magnetic states of actinides. Structures were curated to remove non-converged geometries and spurious high-energy outliers.

• Model Training (MACE-Osaka26):

  • Integration: HE26 was integrated with existing molecular (OFF23) and crystalline (MPtrj) datasets using the Total Energy Alignment (TEA) protocol to ensure energy consistency across domains.
  • Architecture: The model is based on the MACE (Higher-order Equivariant Message Passing) framework.
  • Key Modification: The graph cutoff radius was increased from 4.5 Å (used in previous MACE-Osaka24) to 6.0 Å. This was critical to capture the longer-range interactions essential for heavy element body-centered cubic (BCC) structures and complex oxides.
  • Scope: The resulting model, MACE-Osaka26, covers 97 elements, representing the broadest elemental coverage for a universal MLIP to date.

• Property Prediction:

  • Lattice thermal conductivity ($\kappa_L$) was calculated by solving the phonon Wigner transport equation (WTE) using second- and third-order interatomic force constants (IFCs) derived from the MLIP. This approach captures both particle-like and wave-like (coherent) heat transport, which is vital for low-symmetry disordered systems.

3. Key Contributions

1. HE26 Dataset: The first open-source, high-quality DFT dataset specifically curated for minor actinides and other heavy elements, filling a critical void in materials informatics.
2. MACE-Osaka26 Model: A universal MLIP covering 97 elements. It is the first model capable of handling the full periodic table up to Radon, including the most challenging transuranium elements.
3. Methodological Innovation: Demonstrated that increasing the graph cutoff radius (to 6.0 Å) significantly improves the accuracy of heavy element predictions without sacrificing performance on lighter elements.
4. Thermal Transport in Complex Systems: Successfully applied a universal MLIP to predict lattice thermal conductivity in complex, disordered actinide solid solutions (e.g., Am-doped $UO_2$), a task previously prohibitive for DFT due to the need for massive supercells.

4. Results

• Cross-Domain Accuracy:
  • Crystalline (MPtrj): Improved energy RMSE from 77.2 to 62.0 meV/atom compared to the previous model.
  • Molecular (OFF23): Significantly enhanced energy accuracy, halving the RMSE from 25.2 to 12.3 meV/atom, proving robust generalization to isolated molecules.
  • Heavy Elements (HE26): Achieved an energy MAE of 44.7 meV/atom and RMSE of 117.3 meV/atom on the training set. Notably, force and stress accuracies were high (Force MAE: 26.3 meV/Å), indicating the model learned complex potential energy surfaces despite limited data.
• Lattice Parameter Prediction:
  • For the BHE subset, the model achieved an RMSE of ~0.45 Å, with near-zero errors for Am and Cm.
  • For the CHE subset (multi-component), the RMSE dropped to 0.0477 Å, demonstrating excellent generalization to complex chemical environments.
  • For CCFO (fluorite oxides), the model reproduced DFT trends across the actinide series (Th to Cf) with an RMSE of 0.0067 Å.
• Thermal Conductivity ($\kappa_L$):
  • The model accurately predicted $\kappa_L$ for actinide dioxides ($AnO_2$) across a wide temperature range (300–1500 K), matching experimental data for $UO_2$, $ThO_2$, and others.
  • Crucially, the model predicted no imaginary phonon frequencies for any of the eight actinide systems, confirming dynamical stability.
  • For the mixed oxide solid solution $Am_{0.25}U_{0.75}O_2$, the model quantitatively captured the severe degradation of thermal conductivity caused by Am doping, bridging the properties of the pure end-members.

5. Significance

• Accelerating Nuclear Material Design: This work provides a scalable, computationally efficient route to screen and design nuclear fuels, waste forms, and high-entropy ceramics containing transuranium elements, which was previously impossible using DFT alone due to computational costs.
• Foundation for Heavy Element Science: By opening the chemical space to 97 elements, MACE-Osaka26 serves as a foundational tool for materials informatics, enabling the prediction of properties beyond energy (e.g., elasticity, thermal transport) for the entire periodic table.
• Open Science: Both the HE26 dataset and the MACE-Osaka26 model are publicly available, establishing a reproducible starting point for future research into heavy element physics and chemistry.
• Future Outlook: The authors note that while the current model uses PBE, future iterations should incorporate Spin-Orbit Coupling (SOC) and higher-level electronic structure methods to further refine predictions for strongly correlated systems.