Harvard University Press

This work applies scientific multi-agent reinforcement learning (SMARL) to subgrid-scale closure modeling of geophysical turbulence: using the enstrophy spectrum estimated from a few high-fidelity samples as reward, it learns Leith model coefficients online, enabling LES with 160 to 163,840 times coarser resolution than DNS to run stably for simulations about 2000 times the length of the training data and to reproduce DNS kinetic energy spectra and vorticity probability density functions, including the tails that represent extreme events.

This work applies scientific multi-agent reinforcement learning (SMARL) to subgrid-scale closure modeling of geophysical turbulence: using the enstrophy spectrum estimated from a few high-fidelity samples as reward, it learns Leith model coefficients online, enabling LES with 160 to 163,840 times coarser resolution than DNS to run stably for simulations about 2000 times the length of the training data and to reproduce DNS kinetic energy spectra and vorticity probability density functions, including the tails that represent extreme events.

This work proposes a new self-supervised pretraining task for 3D graph neural networks: pretrained on 542k SwissProt protein structures from the AlphaFold Database, the model predicts the Euclidean distance between the geometric centroid of each protein subgraph (2-hop ego networks centered on 10% of amino acids) and the global geometric centroid of the whole protein, discretized into 10 equal bins and trained with cross-entropy; across ProNet, SchNet, and GCN backbones and ca_base, ca_angles, and ca_bb featurizations, it improves Fold, Superfamily, Family, and React classification by up to about 6% over no-pretraining and edge-distance pretraining baselines, without multiple views, augmentations, or masking strategies.

This work proposes a new self-supervised pretraining task for 3D graph neural networks: pretrained on 542k SwissProt protein structures from the AlphaFold Database, the model predicts the Euclidean distance between the geometric centroid of each protein subgraph (2-hop ego networks centered on 10% of amino acids) and the global geometric centroid of the whole protein, discretized into 10 equal bins and trained with cross-entropy; across ProNet, SchNet, and GCN backbones and ca_base, ca_angles, and ca_bb featurizations, it improves Fold, Superfamily, Family, and React classification by up to about 6% over no-pretraining and edge-distance pretraining baselines, without multiple views, augmentations, or masking strategies.