PET-MAD, a universal interatomic potential for advanced materials modeling (2503.14118v1)

Abstract: Machine-learning interatomic potentials (MLIPs) have greatly extended the reach of atomic-scale simulations, offering the accuracy of first-principles calculations at a fraction of the effort. Leveraging large quantum mechanical databases and expressive architectures, recent "universal" models deliver qualitative accuracy across the periodic table but are often biased toward low-energy configurations. We introduce PET-MAD, a generally applicable MLIP trained on a dataset combining stable inorganic and organic solids, systematically modified to enhance atomic diversity. Using a moderate but highly-consistent level of electronic-structure theory, we assess PET-MAD's accuracy on established benchmarks and advanced simulations of six materials. PET-MAD rivals state-of-the-art MLIPs for inorganic solids, while also being reliable for molecules, organic materials, and surfaces. It is stable and fast, enabling, out-of-the-box, the near-quantitative study of thermal and quantum mechanical fluctuations, functional properties, and phase transitions. It can be efficiently fine-tuned to deliver full quantum mechanical accuracy with a minimal number of targeted calculations.

Overview of PET-MAD: A Universal Interatomic Potential for Advanced Materials Modeling

The paper introduces PET-MAD, a machine-learning interatomic potential (MLIP) designed for advanced materials modeling, particularly focusing on both inorganic and organic solids. The starting point for PET-MAD is the recognition of the need for a universal model that can achieve the accuracy of first-principles calculations efficiently across the periodic table. While PET-MAD achieves impressive accuracy on established benchmarks for inorganic solids, its reliability extends to molecules, organic materials, and surfaces due to its comprehensive training dataset and robust architectural design.

Key Contributions

1. Massive Atomistic Diversity (MAD) Dataset:
   • The MAD dataset incorporates a high degree of chemical and structural diversity essential for training a universal model. This dataset combines a broad spectrum of stable inorganic and organic solids, systematically modified to enhance atomic diversity and include non-equilibrium structures.
   • Calculations for the dataset utilize a consistent electron density functional theory, allowing a quantitative assessment across different chemical systems. This design choice ensures coherence in potential learning across the periodic table, albeit with some intrinsic limitations.
2. Point Edge Transformer (PET) Architecture:
   • The PET architecture is leveraged for its flexibility and expressiveness, due to its transformer-based graph neural network (GNN) structure. This allows the model to approximate energy and force predictions accurately without imposing explicit rotational symmetry constraints.
   • By utilizing a low-rank adaptation (LoRA) technique for fine-tuning, PET-MAD is capable of maintaining precision while adapting to specific system requirements with minimal calculations.
3. Performance and Benchmarking:
   • PET-MAD's accuracy is critically assessed against state-of-the-art MLIPs: MACE-MP-0-L, MatterSim-5M, Orb-v2, and SevenNet-l3i5. It demonstrates superior performance in various datasets indicating its robustness for diverse chemical systems.
   • Its speed and memory efficiency during simulations further position PET-MAD as a highly accessible tool for complex atomistic simulations.
4. Application in Advanced Simulations:
   • Six illustrative examples demonstrate PET-MAD’s reliability in predicting critical functional properties and showcasing advanced simulation capabilities, such as ionic transport in lithium thiophosphate and quantum mechanical fluctuations in liquid water.

Implications for Materials Science

The development and implementation of PET-MAD hold significant implications for both theoretical and practical advancements in materials science:

• Versatility: The universal applicability across various chemical systems makes PET-MAD an invaluable tool for exploratory research and development of novel materials.
• Efficiency: By offering near-quantitative predictions out of the box, PET-MAD reduces the need for extensive calculations that bespoke models require, accelerating the research cycle for new materials.
• Future Developments: The methodological advances demonstrated by PET-MAD pave the way for further integration of machine learning in material design and discovery, emphasizing the need for efficient, large-scale dataset preparation and robust transformation architectures.

Overall, PET-MAD represents a potent advancement in the endeavor to transcend traditional computational limitations and adopt AI-driven methods into materials modeling, suggesting its potential as a staple tool in future AI-based material science research initiatives.