Mixture of Experts for Interatomic Potentials in the NequIP and Allegro Framework

Hybrid atomistic simulation strategies that combine different levels of accuracy have long been employed in QM/MM methods and are now emerging within the landscape of machine-learning interatomic potentials (MLIPs). Such approaches promise scalable and efficient simulations of heterogeneous systems containing both simple and complex regions. In this work, we introduce a mixture-of-experts framework for the NequIP and Allegro graph neural network architectures, extending their state-of-the-art accuracy to mixed-fidelity molecular dynamics. We develop co-training strategies that enable multiple expert models to jointly learn a unified potential energy surface while balancing specialization in distinct regions and consistency at their interfaces. A new interface within LAMMPS allows for efficient, energy-conserving multi-model inference, preserving differentiability and parallel performance. The framework further integrates with NequIP and Allegro foundation models, enabling fine-tuning and co-training workflows that accelerate the construction of mixed-fidelity models from pretrained experts. Together, these advances provide yet another step in the magnitude of system sizes and timescales reachable by these methods.