DP-EVA: data-efficient fine-tuning framework via maximizing pretrained knowledge of large atomistic models for developing domain-specific machine-learning interatomic potentials
Clean Energy · 09.07.2026 · 10 (4) · 103-116
DOI: 10.1093/ce/zkag029 ↗Kurz gefasst
Fachartikel in der Zeitschrift Clean Energy, online erschienen am 11.06.2026. Verfasst von Zhaoqing Liu und 5 weiteren Autor:innen. Die Arbeit nennt 94 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Machine-learning interatomic potentials (MLIPs) have greatly extended the temporal and spatial scales of atomistic simulations, enabling the theoretical study of complex processes at affordable computational cost compared with conventional density functional theory. Recently, large atomistic models (LAMs) have drawn intense interest, as their unified encoders embed extensive chemical knowledge and support fine-tuning methodologies for efficiently adapting models to domain-specific downstream tasks. While many active learning frameworks exist for building MLIP datasets from scratch, dedicated data-generation pipelines for fine-tuning pretrained LAMs remain scarce. Here, we introduce the Deep Potential EVolution Accelerator (DP-EVA)—a data-efficient fine-tuning framework that maximizes the utilization of the pretrained knowledge of LAMs during data generation, accelerating the evolution of domain-specific fine-tuned models with minimal datasets. DP-EVA collects highly representative data through a dual-dimensional shallow-ensemble-based uncertainty quantification method based on parallel fine-tuning on the LAM decoder and a DImensionality-Reduced Encoded Clusters with sTratified (DIRECT) sampling strategy based on the LAM encoder. Tests show that DP-EVA delivers optimal chemical space coverage in the task of drastically reducing the size of an existing MLIP dataset for iron-based Fischer–Tropsch synthesis. DP-EVA fills the gap of active learning frameworks suitable for fine-tuning LAMs toward domain-specific MLIPs and it is also open-source, Slurm-native, and agent-ready for the coming era of agentic scientific research.
Zitieren
APA
Liu, Z., Deng, Z., Zhao, H., Wang, H., Chen, M., & Jiang, H. (2026). DP-EVA: data-efficient fine-tuning framework via maximizing pretrained knowledge of large atomistic models for developing domain-specific machine-learning interatomic potentials. Clean Energy, 10(4), 103-116. https://doi.org/10.1093/ce/zkag029
Vancouver
Liu Z, Deng Z, Zhao H, Wang H, Chen M, Jiang H. DP-EVA: data-efficient fine-tuning framework via maximizing pretrained knowledge of large atomistic models for developing domain-specific machine-learning interatomic potentials. Clean Energy. 2026;10(4):103-116. doi:10.1093/ce/zkag029
BibTeX
@article{Liu2026,
author = {Liu, Zhaoqing and Deng, Zhe and Zhao, Huabo and Wang, Han and Chen, Mohan and Jiang, Hong},
title = {DP-EVA: data-efficient fine-tuning framework via maximizing pretrained knowledge of large atomistic models for developing domain-specific machine-learning interatomic potentials},
journal = {Clean Energy},
year = {2026},
volume = {10},
number = {4},
pages = {103--116},
doi = {10.1093/ce/zkag029}
}
EUVIMED