.png){kind=icon}
Return
Cluster-based Bayesian optimization for application-specific interatomic potential training
.png){kind=icon}
PRE
AI
DOI：10.1016/j.commatsci.2026.114943.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
While significant strides have been made to develop universal (foundation) interatomic potentials, these models are often too computationally expensive for large-scale molecular dynamics or lack the necessary accuracy for certain applications of interest due to the use of overly-general training sets. Because of this, there is a persistent need for “bespoke” potential fitting — training fast, specialized models on a case-by-case basis that target specific regions of chemical space. In this work, we develop a multi-level optimization algorithm that uses cluster-based training to efficiently learn training data weights that are optimally informative for the properties of interest, while simultaneously optimizing the model hyperparameters and maintaining transferability. We demonstrate this approach by training a bespoke SNAP model for carbon beginning with an existing carbon dataset and show an order of magnitude improvement in prediction of force constants and radial distribution functions when the training set is optimally weighted to select the most informative training points, compared to the baseline model trained on the full dataset. We compare the performance of HDBSCAN and k -means for clustering with bispectrum descriptors, autoencoder embeddings, and pre-trained interatomic potential embeddings, and find that HDBSCAN performs well when used with atomic structure descriptors that closely match those used during inference (bispectrum or pre-trained model embeddings). Finally, we demonstrate the viability of this approach in real-world fitting scenarios by using it to fit a potential that balances accurate prediction of force constants, radial distribution function, and training set force mean absolute errors.
Keywords:
Interatomic potential
Fine-tuning
Clustering
Journal
IF:
3.3
Papers:
1.3W
Citations:
3.6W
Organization
Cited Papers
No cited papers available