OpenNovelty.org

DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
machine learning interatomic potentialmolecular dynamicsatomistic simulation
Abstract:

Large-scale atomistic simulations are essential to bridge computational materials and chemistry to realistic materials and drug discovery applications. In the past few years, rapid developments of machine learning interatomic potentials (MLIPs) have offered a solution to scale up quantum mechanical calculations. Parallelizing these interatomic potentials across multiple devices poses a challenging, but promising approach to further extending simulation scales to real-world applications. In this work, we present \textbf{DistMLIP}, an efficient distributed inference platform for MLIPs based on zero-redundancy, graph-level parallelization. In contrast to conventional space-partitioning parallelization, DistMLIP enables efficient MLIP parallelization through graph partitioning, allowing multi-device inference on flexible MLIP model architectures like multi-layer graph neural networks. DistMLIP presents an easy-to-use, flexible, plug-in interface that enables distributed inference of pre-existing MLIPs. We demonstrate DistMLIP on four widely used and state-of-the-art MLIPs: CHGNet, MACE, TensorNet, and eSEN. We show that existing foundation potentials can perform near-million-atom calculations at the scale of a few seconds on 8 GPUs with DistMLIP.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces DistMLIP, a distributed inference platform for machine learning interatomic potentials that employs graph-level parallelization rather than conventional space decomposition. It resides in the 'Graph-Level Parallelization Platforms' leaf of the taxonomy, which contains only two papers total. This sparse population suggests the research direction is relatively nascent, with limited prior work explicitly focused on graph partitioning strategies for MLIP inference. The taxonomy indicates that distributed inference frameworks as a whole remain an emerging area within the broader MLIP ecosystem.

The taxonomy tree reveals that DistMLIP's parent branch, 'Distributed and Parallel Inference Frameworks', includes neighboring leaves for multi-node training systems and foundation model optimization. These adjacent directions address scalability through different lenses: training-time parallelism versus inference-time partitioning, or architectural pruning versus runtime distribution. The scope notes clarify that space-decomposition methods and training-focused parallelization belong elsewhere, positioning DistMLIP's graph-level approach as a distinct alternative to domain decomposition techniques commonly used in classical molecular dynamics. This structural context highlights how the work diverges from both traditional spatial partitioning and training-centric distributed systems.

Among the three contributions analyzed, the core platform concept examined ten candidates and found one potentially refutable prior work, suggesting moderate overlap in the limited search scope. The graph-level partitioning method and plug-in interface contributions each examined five to six candidates with no clear refutations, indicating these aspects may be more novel within the twenty-one papers reviewed. The statistics reflect a focused semantic search rather than exhaustive coverage, so the absence of refutations for two contributions does not guarantee absolute novelty but does suggest these elements are less directly addressed in the immediate literature neighborhood.

Based on the limited search scope of twenty-one candidates, the work appears to occupy a relatively underexplored niche within MLIP parallelization. The sparse taxonomy leaf and contribution-level statistics suggest that graph partitioning for MLIP inference has received less attention than training workflows or domain-specific applications. However, the analysis does not cover the full breadth of parallel computing or molecular dynamics literature, leaving open the possibility of relevant work outside the top-K semantic matches examined here.

Taxonomy

Core-task Taxonomy Papers
16
3
Claimed Contributions
21
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: distributed inference for machine learning interatomic potentials. The field encompasses a diverse set of approaches for deploying and applying machine learning interatomic potentials (MLIPs) at scale. The taxonomy reveals several major branches: Distributed and Parallel Inference Frameworks focus on computational strategies for efficient evaluation of potentials across many atoms or configurations, often leveraging graph-level or domain decomposition parallelism. Active Learning and Training Data Generation addresses the iterative refinement of training sets to improve model accuracy with minimal computational cost. Domain-Specific MLIP Applications demonstrate the use of these potentials in targeted materials contexts, such as battery cathodes, alloys, and nanostructures. Software Packages and Implementation Tools provide the practical infrastructure for researchers to build and deploy MLIPs, while Multiscale and Hybrid Simulation Frameworks integrate MLIPs with coarser or finer-scale methods, and Classical Force Field Optimization explores traditional parameterization techniques that complement or compete with machine learning approaches. Within the distributed inference landscape, a particularly active line of work centers on graph-level parallelization platforms, where the challenge is to partition large atomic graphs efficiently for concurrent evaluation. DistMLIP[0] sits squarely in this branch, addressing scalability for graph neural network potentials alongside closely related efforts such as Scalable GNN Potentials[3], which similarly targets efficient parallel inference for equivariant architectures. Nearby works like High Performance Equivariant Potentials[1] and aims-PAX Parallel[9] explore complementary strategies for accelerating inference, whether through optimized kernels or hybrid parallelization schemes. A key trade-off across these studies is balancing communication overhead against computational load, especially as system sizes grow. DistMLIP[0] emphasizes graph partitioning and distributed memory strategies, contrasting with approaches that rely more heavily on shared-memory or GPU-centric optimizations, thus offering a distinct perspective on how to scale MLIP inference to very large simulations.

Claimed Contributions

DistMLIP: A distributed inference platform for MLIPs using graph-level parallelization

The authors introduce DistMLIP, a platform that enables multi-device inference of machine learning interatomic potentials through graph partitioning rather than spatial partitioning. This approach achieves zero redundancy by avoiding redundant computation on ghost atoms and supports flexible MLIP architectures including multi-layer graph neural networks.

10 retrieved papers
Can Refute
Graph-level partitioning method for distributing atom and three-body bond graphs

The authors develop a graph partitioning technique that distributes both atom graphs and augmented three-body line graphs across multiple devices. This method enables efficient parallelization of long-range GNN-based MLIPs by transferring node and edge features between partitions at each convolution layer while preserving gradient computation capability.

6 retrieved papers
Plug-in interface for flexible distributed inference of pre-existing MLIPs

The authors provide a standalone, model-agnostic interface that does not depend on third-party distributed simulation libraries like LAMMPS. This design allows most popular MLIPs to be adapted with minimal modification and supports flexible usage across different MLIP workflows.

5 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

DistMLIP: A distributed inference platform for MLIPs using graph-level parallelization

The authors introduce DistMLIP, a platform that enables multi-device inference of machine learning interatomic potentials through graph partitioning rather than spatial partitioning. This approach achieves zero redundancy by avoiding redundant computation on ghost atoms and supports flexible MLIP architectures including multi-layer graph neural networks.

Contribution

Graph-level partitioning method for distributing atom and three-body bond graphs

The authors develop a graph partitioning technique that distributes both atom graphs and augmented three-body line graphs across multiple devices. This method enables efficient parallelization of long-range GNN-based MLIPs by transferring node and edge features between partitions at each convolution layer while preserving gradient computation capability.

Contribution

Plug-in interface for flexible distributed inference of pre-existing MLIPs

The authors provide a standalone, model-agnostic interface that does not depend on third-party distributed simulation libraries like LAMMPS. This design allows most popular MLIPs to be adapted with minimal modification and supports flexible usage across different MLIP workflows.

Table of Contents

DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials | Novelty Validation