NEP89 enables universal neuroevolution for 89 inorganic and organic elements

In a transformative advancement for materials science, researchers have unveiled NEP89, a universal neuroevolution potential model that bridges the long-standing gap between computational efficiency and accuracy in atomistic simulations. Unlike prior machine-learned interatomic potentials, which often suffer from either computational heaviness or limited material scope, NEP89 spans an impressive 89 elements, covering both inorganic and organic materials with remarkable precision.

Atomistic simulations traditionally rely on interatomic potentials to model the forces between atoms, providing insights into material properties at the quantum level. However, the most accurate potentials, often derived from quantum mechanical calculations, are prohibitively computationally expensive for large systems. Conversely, faster empirical potentials tend to lack the precision needed for complex or novel materials. NEP89, built upon the neuroevolution potential (NEP) architecture, strikes a new balance by delivering near-empirical-potential speed while maintaining nearly quantum-level accuracy.

This breakthrough was enabled through the curation of an extensive yet compact training dataset, achieved by an innovative descriptor-space subsampling strategy. Iterative refinement across multiple datasets ensured that the model’s training encompassed diverse material chemistries and configurations, thereby enhancing its universal applicability. The resulting model does not merely replicate accuracy benchmarks; it does so while being three to four orders of magnitude faster than existing foundation models, making it feasible to simulate previously unattainable large-scale systems at the atomic level.

NEP89’s versatility shines in a range of complex simulation scenarios. It has been successfully applied to million-atom compressions of compositionally complex alloys, revealing insights unattainable before due to computational constraints. Its utility extends to modeling ion diffusion in solid-state electrolytes and water, processes critical to battery and fuel cell technology. Additionally, NEP89 effectively captures phenomena such as rocksalt dissolution, methane combustion, and intricate protein-ligand interactions, demonstrating its potential to impact fields as diverse as catalysis and biochemistry.

Beyond its out-of-the-box capabilities, NEP89 supports fine-tuning for specialized applications, allowing researchers to tailor the potential for targeted studies. This adaptability is vital for probing mechanical, thermal, structural, and spectral properties in challenging material classes such as two-dimensional materials, metallic glasses, and organic crystals. As a result, NEP89 provides a flexible platform capable of addressing evolving research demands in materials science.

The advancement heralded by NEP89 offers a significant step toward universal and scalable atomistic modeling. By marrying speed and accuracy on a broad elemental palette, it paves the way for unprecedented exploration of material behaviors at scales and complexities previously beyond reach, potentially accelerating innovations in materials design and discovery.

This breakthrough underscores the growing role of machine learning and neuroevolution techniques in computational materials science, emphasizing how strategic data curation and algorithmic innovation can overcome longstanding trade-offs. As NEP89 becomes integrated into simulation toolkits, its impact will resonate across disciplines that rely on atomic-scale insights to drive technological progress.

With this development, NEP89 not only embodies a new benchmark in performance but also sets a visionary standard for the next generation of machine-learned potentials, promising to reshape the landscape of computational materials research.