Active Learning of Uniformly Accurate Inter-atomic Potentials for Materials Simulation (1810.11890v2)

Abstract: An active learning procedure called Deep Potential Generator (DP-GEN) is proposed for the construction of accurate and transferable machine learning-based models of the potential energy surface (PES) for the molecular modeling of materials. This procedure consists of three main components: exploration, generation of accurate reference data, and training. Application to the sample systems of Al, Mg and Al-Mg alloys demonstrates that DP-GEN can produce uniformly accurate PES models with a minimal number of reference data.

• The paper develops DP-GEN, an active learning framework that reduces computational cost by labeling only the most informative configurations.
• It employs a deep neural network-based DP model to capture complex many-body interactions with inherent physical symmetries.
• The method outperforms traditional force fields, demonstrating superior accuracy across Al, Mg, and Al-Mg alloy systems.

Active Learning of Uniformly Accurate Inter-atomic Potentials for Materials Simulation

The paper entitled "Active Learning of Uniformly Accurate Inter-atomic Potentials for Materials Simulation" introduces an active learning framework named Deep Potential Generator (DP-GEN) for constructing highly accurate machine learning models of the potential energy surface (PES) for materials modeling. This DP-GEN approach aims to balance computational efficiency and accuracy by reducing the reliance on extensive and computationally expensive {\it ab initio} data, thereby rendering simulations more feasible at larger scales and longer timescales than traditional methods.

Core Contributions

1. Active Learning Framework: The DP-GEN scheme is modular, comprising exploration, data labeling, and model training components. This modular approach facilitates flexibility and potential customization for various systems. It aims at efficiently sampling the configuration space and strategically selecting data points where model predictions are less reliable for labeling through high-level quantum calculations.
2. DP Model: The Deep Potential (DP) model is employed for representing the PES. It is characterized by its capacity to capture complex many-body interactions while preserving symmetries intrinsic to physical systems, such as translational and rotational invariance. The potential is expressed as a sum over atomic contributions, with each atomic energy function being a deep neural network parameterized to take the local atomic environment into account.
3. Sampling and Testing: For the structural testing, DP-GEN was applied to Al, Mg, and Al-Mg alloy systems to create a uniformly accurate PES that adequately describes these materials over various compositions and thermodynamic states. The innovation of DP-GEN lies not only in its predictive ability but also its substantial reduction in required computational resources compared to traditional methods like AIMD.

Strong Numerical Results

• Efficiency and Accuracy: DP-GEN demonstrated its efficacy by exploring upwards of 650 million configurations but labeled only a minuscule fraction (0.0044%), showcasing its efficiency in identifying the most informative data points.
• Comparative Analysis: The comparison against empirical force fields, such as the modified embedded-atom method (MEAM), and density functional theory (DFT) illustrates superior performance by the DP model in accurately predicting properties like phonon spectra, elastic constants, and others, reducing errors associated with FFs.
• Multi-scale Capabilities: Notably, DP simulations could test system properties under high-throughput schemes in Mg-Al alloys, applying the model successfully to larger periodic structures than initially trained upon.

Theoretical and Practical Implications

This research provides a significant advancement in computational materials science. By minimizing necessary {\it ab initio} data and leveraging active learning, DP-GEN enhances the feasibility of constructing PES models applicable to diverse materials, including those with metallic bonding, potentially extendable to ceramics and polymers. Its adaptability in configuration sampling and capacity for automation highlight the practical utility of DP-GEN in expediting simulations for material discovery and innovation.

Future Prospects and Considerations

The flexible structure of DP-GEN enables future developments where sampling strategies or enhanced learning algorithms, perhaps using metadynamics or genetic algorithms, could be incorporated seamlessly. It challenges the convention of hypothesis-driven empirical FF development, leaning towards data-driven representations powered by machine learning. However, challenges remain, especially in handling high-dimensional phase spaces or ensuring the robustness of the uncertainty indicator in extreme cases. Ensuring reliable performance in new systems necessitates ongoing validation, yet the ability to construct models from scratch promises new frontiers in computational efficiency and scalability.

In conclusion, the paper showcases a substantial methodological advancement in machine learning-enhanced materials simulation. DP-GEN achieves a balance between accuracy and efficiency that could be pivotal in tackling broader classes of materials, thus expanding the horizon of computational materials research.