APO: Unsupervised Atomic Policy Optimization for 3D Structure Prediction of Atomic Systems

Imagine you're working on developing new materials or drugs, and you need to predict the 3D structures of atomic systems. Traditionally, this involves using supervised learning methods that require a lot of labeled data, which can be hard to come by, especially for novel materials or proteins. This reliance on ground-truth coordinates can create a bottleneck, as obtaining experimental labels is often prohibitively expensive. This is what's called the data-scarcity problem, where the lack of sufficient labeled data limits the effectiveness of the models.

To address this, a new approach called Atomic Policy Optimization (APO) has been proposed. Instead of needing labeled data, APO uses a fully unsupervised framework that aligns atomic structures based on their physical properties. It employs a dual-reward mechanism that encourages the model to find stable and plausible configurations without needing external labels. By focusing on intrinsic physical consistency, APO not only improves the accuracy of structure predictions but also enhances inference efficiency. Compared to previous methods, this approach allows for better performance in predicting structures, making it a valuable tool for researchers in material science and drug discovery.

The approach introduces a fully unsupervised method for 3D structure prediction, which is a significant shift from reliance on supervised learning.

The benchmarks against fully supervised baselines are extensive, demonstrating solid performance improvements.