Abstract:Machine-learned interatomic models have growing in popularity due to their ability to afford near quantum-accurate predictions for complex phenomena, with orders-of-magnitude greater computational efficiency. However, these models struggle when applied to systems of many element types due to the near exponential increase in number of parameters that must be determined. To mitigate this challenge, we present a new hierarchical transfer learning approach that allows the fitting problem to be decomposed into smaller independent and reusable parameter blocks that enable development of explicitly chemically extensible ML- IAM. Application of this strategy is demonstrated for C and N mixtures under conditions ranging from nominally ambient to approximately 10,000 K and 200 GPa, and compositions from 0 to 100 % N. Ultimately, this strategy makes model generation for chemically complex systems more tractable and efficient, facilitates comprehensive model validation, and makes ML-IAM development for problems of this nature more accessible to users with limited access to extreme computing infrastructure.

What problem does this paper attempt to address?

This paper mainly discusses the challenges faced by machine learning (ML) in constructing atomic models, especially the problem of a sharp increase in the number of parameters when multiple types of elements are involved. To address this issue, the authors propose a new hierarchical transfer learning strategy that decomposes the parameter fitting problem into independent and reusable blocks, enabling the development of chemically scalable ML-IAMs (machine learning atomic models). This approach makes model generation more feasible and efficient for chemically complex systems, and facilitates comprehensive model validation, making it easier for resource-limited users to study such problems. In the introduction, the authors mention their use of a ML-IAM called ChIMES and demonstrate its application in carbon (C) and nitrogen (N) mixtures under different temperature and pressure conditions. They evaluate the accuracy of the hierarchical transfer learning model by comparing its performance with traditional direct learning methods. The research shows that high-quality C/N models can be constructed even without re-fitting the parameters of pure C and N systems. In addition, the paper discusses the advantages of this approach, including reducing model complexity, better defining the physical and chemical conditions applicable to the model, and improving model transferability and generation efficiency. Through hierarchy, it can be immediately applied to subsets of specific chemical spaces while gradually improving and expanding the scope of the model to achieve real-time problem-solving. In the method and computational details section, the authors provide a detailed description of the architecture of the ChIMES model, descriptors, and how parameters are optimized through hierarchy and transfer learning. They compare the performance of the transfer learning model with density functional theory (DFT) results by simulating the performance of C/N mixtures under various conditions, proving the performance and efficiency of the transfer learning model. In summary, this paper proposes a new hierarchical transfer learning strategy to address the complexity and efficiency issues of machine learning atomic models in dealing with multi-element systems, providing a more effective method for simulating chemically complex systems.

Hierarchical Transfer Learning: An Agile and Equitable Strategy for Machine-Learning Interatomic Models

Transferable Prediction of Intermolecular Coupling Achieved by Hierarchical Material Representation

Exploring the potential of transfer learning for metamodels of heterogeneous material deformation

Representations and Strategies for Transferable Machine Learning Models in Chemical Discovery

Transferability of datasets between Machine-Learning Interaction Potentials

Interpolation and differentiation of alchemical degrees of freedom in machine learning interatomic potentials

Learning Together: Towards foundational models for machine learning interatomic potentials with meta-learning

Towards Efficient, Accurate, Scalable, and Transferable Quantum Machine Learning with AM-ons: the DNA of Chemistry

Accelerating Molecular Dynamics Simulations with Quantum Accuracy by Hierarchical Classification

Training Transferable Interatomic Neural Network Potentials for Reactive Chemistry: Improved Chemical Space Sampling

Learning together: Towards foundation models for machine learning interatomic potentials with meta-learning

Quantum machine learning using atom-in-molecule-based fragments selected on-the-fly

Constructing accurate and efficient general-purpose atomistic machine learning model with transferable accuracy for quantum chemistry

MLatom 2: an Integrative Platform for Atomistic Machine Learning

Leveraging Transfer Learning and Chemical Principles toward Interpretable Materials Properties

Overcoming the Size Limit of First Principles Molecular Dynamics Simulations with an In-Distribution Substructure Embedding Active Learner

Identifying and embedding transferability in data-driven representations of chemical space

Advancing Multiscale Molecular Modeling with Machine Learning-Derived Electrostatics

Towards High-Throughput Microstructure Simulation in Compositionally Complex Alloys Via Machine Learning

Deep machine learning potentials for multicomponent metallic melts: Development, predictability and compositional transferability

Transferability of Atom-Based Neural Networks