Bridging Atomic Simulations and Materials Design through Scalable Strategies for Machine Learning Interatomic Potentials

Machine Learning Interatomic Potentials (MLIPs) are transforming computational materials science by enabling quantum-accurate simulations at unprecedented scales. However, distinct scalability challenges remain: managing increasing training data volumes and achieving massive parallelization for large-scale simulations. Addressing both is essential for extending applicability to complex materials and enabling practical property evaluation for materials design. This work presents three interconnected studies addressing these challenges.

First, we analyze Allegro-FM, a foundation model leveraging E(3)-equivariant architecture trained on unified organic-inorganic datasets. Its strictly local strategy preserves traditional domain decomposition, enabling billion-atom simulations while facilitating reaction tracking in complex multi-domain structures. We demonstrate these capabilities through a case study examining calcium silicate hydrates and silicate carbonation at solid-liquid interfaces.

Second, building on foundation model scalability, we examine our Edge-wise Emergent Decomposition (E3D) framework, which analyzes how neural networks learn chemical properties. By decomposing edge-wise energy contributions in trained Allegro models, E3D reveals that networks implicitly learn bond-specific energies correlating with experimental bond dissociation energies. We present additional metrics for model reliability and training convergence assessment.

Additionally, we also present thermal transport simulations using lightweight Atomic Cluster Expansion (ACE) models for 100-500nm silicon thin films as a practical demonstration of scalable MLIPs. This capability proves crucial for atomic-level device property evaluation where both accuracy and system size must be satisfied.