BranchIP learns adaptive tensor-product computation and speeds equivariant MLIPs by up to 2.4x with up to 2.6x less memory on a heterogeneous catalysis system and a proton-conducting solid acid electrolyte while maintaining physical fidelity

BranchIP is a single-model framework for learned adaptive tensor-product computation, trained with a novel distillation loss. Equivariant machine learning interatomic potentials (MLIPs) have transformed atomistic modeling, but accurate treatment of complex materials and molecular systems demands expensive models, with tensor products a key computational bottleneck, and the recent emergence of foundation-scale MLIPs further exacerbates this challenge; this work learns adaptive computation inside the model rather than relying on a fixed computational depth. Abstract-level evidence: the authors state the framework design and training loss and run experiments on two systems of physical interest; the specific loss form, model-size list, and training details are not given.

On a heterogeneous catalysis system and a proton-conducting solid acid electrolyte, BranchIP accelerates MLIPs across model sizes by up to 2.4x and reduces memory usage by up to 2.6x while maintaining physical fidelity. Relative to existing equivariant MLIP computation paths, the result reports efficiency gains together with physical fidelity within one framework and spans model sizes rather than a single model scale. Abstract-level evidence: upper-bound figures for speedup and memory reduction (2.4x, 2.6x) and two physically motivated test systems are given; per-system numbers, baseline configurations, and error metrics are not provided.

The learned adaptive computation provides model interpretability by revealing which interactions demand deeper computation and showing how computational depth relates to chemical complexity and dynamics. It turns the efficiency mechanism itself into a readable physical signal, making the allocation of computational depth an observable object rather than only an internal implementation detail. Abstract-level evidence: the authors report this interpretability observation in prose; specific interaction cases, complexity measures, or quantitative dynamics relationships are not given.

This work addresses researchers and practitioners who use equivariant machine learning interatomic potentials for atomistic simulation, in settings where computational cost matters for complex materials and molecular systems, particularly the physically motivated cases of a heterogeneous catalysis system and a proton-conducting solid acid electrolyte. Its value is that, while maintaining physical fidelity, it turns the computational depth of the tensor-product bottleneck into a learnable, allocatable object, yielding efficiency gains across model sizes within a single model and offering a path to relieve the cost pressure of foundation-scale MLIPs; the learned computational depth can also serve as a window onto how interactions relate to chemical complexity and dynamics for follow-up mechanistic study and model diagnosis.

Readers will still want to know: the specific form of the novel distillation loss and how training is set up; the itemized speedup and memory figures per system and per model size together with baseline comparisons; which metrics measure physical fidelity and at what error magnitude; the criteria by which an interaction is judged to demand deeper computation, and how quantitatively computational depth is tied to chemical complexity and dynamics. Because this assessment rests on the abstract only, without the body figures and experiment tables, these details cannot be confirmed from the available text and remain open questions to check against the original.