MACE-Osaka26: A Universal Machine Learning Interatomic Potential Model Covering 97 Elements

MACE-Osaka26 is an open-source Machine Learning Interatomic Potential (MLIP) suite developed by the Osaka University team. As an upgraded version of MACE-Osaka24, its core highlight is supporting 97 elements (including rare earth and actinide elements). Based on the MACE architecture and total energy alignment technology, it combines DFT-level accuracy with efficient computing capabilities, which can solve the computational bottleneck of traditional material simulation and provide a universal tool for energy material design, complex compound research, and other fields.

Traditional Density Functional Theory (DFT) has high accuracy but is computationally expensive, making it difficult to handle large-scale systems or long-time dynamic simulations. Machine Learning Interatomic Potentials (MLIP) achieve efficient simulations with near-first-principles accuracy by learning DFT data, but early MLIP models were limited to specific elements or material types and lacked universality.

MACE-Osaka26 is based on the MACE (Equivariant Message Passing Neural Network) architecture, which strictly maintains physical symmetries such as translation, rotation, and permutation invariance. The core innovation is the total energy alignment technology, which can integrate multi-source data, eliminate systematic biases, improve cross-domain transfer capabilities, and achieve robust prediction of multiple types of materials.

The key breakthroughs of MACE-Osaka26 include:

1. Covering 97 elements, including rare earth and actinide elements (such as uranium and plutonium);
2. Introducing the HE26 heavy element training dataset to enhance the ability to describe complex compounds;
3. Compared with MACE-Osaka24, the element coverage has expanded from mainly light elements to 97 elements, with more complete actinide support, and the applicable fields have extended to strategic areas such as nuclear materials and rare earth materials.

The model is applicable to multiple scenarios:

• Crystalline materials: Predicting lattice constants, elastic moduli, etc. of metals, semiconductors, oxides, etc.;
• Molecular systems: Studying isolated molecules, gas-phase reactions, and interface processes (such as catalysis and adsorption);
• Energy materials: Simulation of nuclear fuel radiation damage, ion transport in battery materials, design of rare earth catalysts, etc.

MACE-Osaka26 provides pre-trained models. The typical workflow is: load the model → prepare input structure → run simulation → analyze results. Hardware requirements: Windows 10+, 8GB+ memory, 500MB+ space, 64-bit processor. It supports users to fine-tune with private data or train custom models from scratch.

MACE-Osaka26 breaks the application boundaries of traditional MLIP, lowers the threshold for material research, accelerates candidate material screening and experimental guidance, and helps explore complex systems such as nuclear materials. In the future, it will play an important role in cutting-edge fields such as the Materials Genome Initiative and AI-driven material design.