Machine-learning force fields have shown great promise in enabling more accurate molecular dynamics simulations than traditional, manually generated force fields. Recent advances have been achieved by leveraging prior knowledge of the physical system, such as rotational, translational, and reflection symmetries. This paper proposes another crucial priori information, previously unexplored: that simulations of molecular systems are inherently continuous, and therefore, continuous states are highly similar. This study demonstrates that this information can be leveraged by restructuring the state-of-the-art equilibrium base model into a Deep Equilibrium Model (DEQ). This approach reuses intermediate neural network features from previous time steps, resulting in 10-20% accuracy and speedup compared to non-DEQ base models on the MD17, MD22, and OC20 200k datasets. Furthermore, training is significantly more memory-efficient, enabling training of more expressive models on larger systems.
