Researchers propose Adaptive Multi-Teacher Routing (ATR) to address the cost and reliability challenges in training universal machine learning interatomic potentials (uMLIPs). ATR reformulates high-fidelity data construction as a structure-wise decision problem, using multiple pretrained uMLIP teachers to estimate prediction reliability.
- The method combines structural descriptors, teacher identity, and inter-teacher disagreement to calibrate teachers against a small set of real r²SCAN labels.
- It selects high-confidence predictions for pseudo-label generation while rejecting structures where no teacher is sufficiently reliable.
- Using only 0.2% of candidate structures for real labels, ATR distills 2.89 million traceable r²SCAN-level pseudo-labels for pretraining.
- A lightweight CHGNet trained on this dataset consistently outperforms baseline and non-routed controls on held-out r²SCAN structures and the MP-r²SCAN benchmark.
- Finite-temperature molecular dynamics shows improved dynamical robustness, maintaining stable trajectories where baseline simulations undergo catastrophic structural collapse.
These results establish active rejection as an effective mechanism for converting multiple pretrained uMLIPs into a scalable and reliable data-construction system for high-fidelity potentials.