Traj2Relax: A Trajectory-Supervised Method for Robust Structure Relaxation.

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Structure relaxation is vital for atomic simulations and materials modeling.
• Traditional methods are computationally expensive, limiting high-throughput applications.

Purpose of the Study:

• Introduce Traj2Relax, a novel trajectory-supervised framework for efficient structure relaxation.
• Enable physically consistent structural convergence across various perturbation magnitudes.

Main Methods:

• Learns a time-dependent velocity field from geometric differences in relaxation trajectories.
• Utilizes conditional velocity field modeling, avoiding explicit energy/force calculations.
• Incorporates a time-scheduled noise mechanism for training stability and deterministic inference.

Main Results:

• Achieves competitive accuracy under near-equilibrium conditions (0.26 Å RMSD, 82.3% space-group consistency).
• Demonstrates superior performance under moderate to strong perturbations (0.38 Å RMSD, 5.8% recovery rate).
• Offers an order-of-magnitude improvement in inference throughput via batch-parallel relaxation.

Conclusions:

• Traj2Relax provides an efficient, physically grounded alternative for learning-driven structure relaxation.
• The framework excels in high-throughput screening, particularly for nonequilibrium or highly perturbed structures.