AlphaNet: Scaling Up Local-frame-based Atomistic Interatomic Potential

TL;DR

AlphaNet tackles the challenge of achieving high accuracy in atomistic simulations at scale by exploiting a local-frame-based $SE(3)$-equivariant architecture that avoids expensive tensor products via frame-based representations and introduces a rotary positional embedding to connect frames across scales. It combines a local-structure encoding with a frame-transition module to capture complex interatomic interactions across metals, covalent, ionic, and long-range bonding, delivering state-of-the-art energy and force predictions across Formate, Defected Graphene, Zeolites, OC20 OC2M, and Matbench Discovery benchmarks. The results show robust scalability with dataset size and system size, with larger training sets yielding substantial reductions in energy and force MAEs while maintaining fast inference on large systems. These findings position AlphaNet as a scalable, transferable neural interatomic potential for multiscale catalysis and materials design, enabling accurate dynamics in large, diverse systems.

Abstract

Molecular dynamics simulations demand an unprecedented combination of accuracy and scalability to tackle grand challenges in catalysis and materials design. To bridge this gap, we present AlphaNet, a local-frame-based equivariant model that simultaneously improves computational efficiency and predictive precision for interatomic interactions. By constructing equivariant local frames with learnable geometric transitions, AlphaNet encodes atomic environments with enhanced representational capacity, achieving state-of-the-art accuracy in energy and force predictions. Extensive benchmarks on large-scale datasets spanning molecular reactions, crystal stability, and surface catalysis (Matbench Discovery and OC2M) demonstrate its superior performance over existing neural network interatomic potentials while ensuring scalability across diverse system sizes with varying types of interatomic interactions. The synergy of accuracy, efficiency, and transferability positions AlphaNet as a transformative tool for modeling multiscale phenomena, decoding dynamics in catalysis and functional interfaces, with direct implications for accelerating the discovery of complex molecular systems and functional materials.

Figures (4)

• Figure 1: Comparisons between AlphaNet predictions and DFT calculations on the Zeolite dataset.
• Figure 2: Mean absolute errors of energy and force predictions of models with $L$ number of layers as a function of training data size on the zeolite dataset.
• Figure 3: Comparisons of inference speed among various NNIP models. (a). The x-axis represents the number of atoms in the system, and the y-axis shows the average time required to predict energy and forces for systems in batches of 10, across a total of 200 batches on the Zeolite dataset. (b). All models are accessible and compliant with the Matbench Discovery leaderboard. The speed is calculated as the number of energy evaluation per second.
• Figure 4: Overview of AlphaNet framework. We first process the input atomic types and coordinates to scalarized features and frames and pass them further to a loop of message passing and frame transition layers, followed by an output block with a temporal connection to the final output. $I/R$ denotes imaginary and real number, $\|\bullet\|$ denotes norm, $\bullet$ denotes element-wise multiplication, and $\circ$ denotes vector scaling.