소재 AI 공유함
Menu
← AI in materials

Materials AI · written by AI

Reducing uncertainty in magnetic interatomic potentials with active learning to lower prediction error: an adaptive spin-lattice learning method

AI editorPublished 2026.10.15

A paper accepted by Physical Review Materials demonstrated more efficient learning by estimating candidate-structure uncertainty for noncollinear spin-lattice potentials with equivariant evidential deep-learning-based active learning. The uncertainty score was strongly correlated with prediction error, and the select-train-fine-tune procedure reduced energy, force, and projected spin-force errors compared with random sampling. [1]

1/ AI 활용 방법: equivariant evidential deep-learning-based active learning
2/ 적용 과제: uncertainty estimation and sample selection for candidate structures in noncollinear spin-lattice interatomic potentials
3/ 보고된 결과: the uncertainty score was strongly correlated with prediction error, and selection, training, and fine-tuning lowered energy, force, and projected spin-force errors compared with random sampling
4/ When training magnetic interatomic potentials, labels that treat noncollinear spin and lattice variables together are expensive, and deciding which structures to compute next is important.
5/ This paper proposed a method for selecting samples by estimating the uncertainty of candidate structures within an active-learning framework using equivariant evidential deep learning.
6/ According to the reported results, the uncertainty score showed a strong correlation with prediction error, and the selection, training, and fine-tuning workflow lowered energy, force, and projected spin-force errors compared with random sampling.
7/ In BiFeO3 and CrTe2, it also showed accuracy comparable to a 4-model committee.
8/ The materials significance of this issue is that it can improve the data efficiency of potentials that include magnetic interactions, showing the potential to improve performance while using fewer high-cost labels.
9/ However, the reported performance is based on specific materials and settings, so the same level of improvement cannot be assumed to be automatically reproducible in other magnetic systems or in wider composition and structure spaces. [1]

How to apply this

To try a procedure that uses uncertainty to more efficiently choose the next structures to calculate when training a magnetic interatomic potential.

  1. Prepare an initial magnetic dataset and train an equivariant interatomic potential that outputs structural prediction uncertainty.
  2. Compute uncertainty scores for candidate structures and prioritize adding the highest-scoring candidates for further calculation.
  3. Retrain or fine-tune the model using the newly obtained labels, and compare validation errors at each iteration.
  4. Compare performance side by side with random sampling or a fixed committee criterion to confirm the actual gain.
Data needed
An initial labeled set including noncollinear spin and lattice degrees of freedom, and training data for energy, force, and spin-related quantities on candidate structures are needed.
Checking the result
Check whether the energy, force, and spin-related errors decrease on a held-out validation set, and inspect the correlation to see whether the uncertainty score rises with the actual error.
What can go wrong
If uncertainty does not faithfully represent actual error, the model may repeatedly choose samples that only appear uncertain or, conversely, miss important regions.

이 가이드는 원문의 방법을 바탕으로 AI가 제안한 적용 방향입니다. 원문에 적힌 내용이 아니며 같은 결과가 재현된다는 보장도 아닙니다.

Related concepts and practice

이 이슈를 읽는 데 필요한 개념입니다. 용어마다 Colab에서 바로 돌려볼 수 있는 실습 노트북이 있습니다.

Sources

  1. Physical Review Materials - Accepted Paper: Tangent-plane evidential uncertainty in active learning for magnetic interatomic potentials

웹 자료를 바탕으로 AI가 자동 요약한 글입니다. 연구의 세부 조건과 결과는 연결된 원문을 확인해주세요.