Atom-JEPA improves molecular and crystal prediction performance through self-supervised pretraining of 3D atomic systems

AI editorPublished 2026.10.06Posted 2026.10.07 18:02 KST

Atom-JEPA was proposed as a joint-embedding predictive architecture for label-free representation learning of 3D atomic structures. After pretraining, the model was transferred to molecular ADMET, quantum chemistry, and crystal property prediction, and it was reported to achieve top performance in ADMET and quantum chemistry while remaining competitive on crystal property prediction. [1]

1/ AI utilization method: self-supervised pretraining based on joint-embedding predictive architecture for 3D atomic systems
2/ Applied task: label-free representation learning of 3D atomic structures followed by transfer to ADMET, quantum chemistry, and crystal property prediction
3/ Reported results: top performance in ADMET and quantum chemistry prediction, and competitive results in crystal property prediction
4/ Atom-JEPA was proposed as a self-supervised pretraining method based on joint-embedding predictive architecture tailored to 3D atomic systems.
5/ It learned representations from unlabeled 3D atomic structures and then transferred them to molecular ADMET prediction, quantum chemistry prediction, and crystal property prediction.
6/ According to the report, this approach showed state-of-the-art performance in ADMET and quantum chemistry prediction, and competitive results in crystal property prediction as well.
7/ In other words, it was presented as an example showing that atomic-level pretrained representations can be reused across different chemistry and materials prediction tasks.
8/ From a materials research perspective, this issue is directly connected to the strategy of building a general-purpose representation from large-scale unlabeled structural data and then improving property prediction with small downstream datasets.
9/ However, these results are limited to specific benchmarks and task settings, and the same level of performance is not automatically guaranteed for other properties or data distributions. [1]

How to apply this

You can create self-supervised pretrained representations from unlabeled 3D atomic structures and then transfer them to small property prediction problems to test performance improvement.

Collect unlabeled 3D atomic structure data and standardize the input representation for pretraining.
Train structural representations with joint-embedding predictive architecture and save the learned model.
Transfer the representation to ADMET, quantum chemistry, and crystal property datasets, and compare against baseline models.
Check performance and error distributions on the test set and through cross-validation to assess generalization.

Data needed: Large-scale 3D atomic structure data and labeled data for downstream evaluation of ADMET, quantum chemistry, and crystal properties are needed.
Checking the result: Compare against the same model trained without pretraining and check whether regression metrics improve consistently on cross-validation or a separate test set.
What can go wrong: Pretraining gains can be overestimated because of differences in structure preprocessing, data splitting, or the distribution of downstream tasks.

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

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

Atom-JEPA improves molecular and crystal prediction performance through self-supervised pretraining of 3D atomic systems

How to apply this

Related concepts and practice

Sources