Reproduce “Multi-Fidelity Training Reshapes Uncertainty Decomposition in Atomistic Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Matteo Tolloso, Davide Bacciu, Luigi Bonati
The authors study multi-fidelity training for molecular-energy prediction on ANI-1x. They pretrain ensembles on approximate DFT energies, then fine-tune on accurate CCSD(T)/CBS energies, using uncertainty to rank structures. Pretraining improves error ranking but can worsen calibration, especially for higher-energy predictions. In one acquisition round, uncertainty-guided selection reduces held-out RMSE more relative to random selection after pretraining; evidence is limited to energy prediction, one estimator, and one setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.