AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

$\texttt{FlashSchNet}$: Fast and Accurate Coarse-Grained Neural Network Molecular Dynamics

Pingzhi Li, Hongxuan Li, Zirui Liu, Xingcheng Lin, Tianlong Chen

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Scalable approximation and equivariant compute

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace CGSchNet with IO-aware fused kernels, CSR segment aggregation and 16-bit quantization (FlashSchNet)CGSchNet · Homeodomain (1ENH), 64 replicasspeed (timestep*mol/s): 477 -> 3095Improved
use FlashSchNet's IO-aware design to avoid materializing edge tensorsCGSchNet · Homeodomain (1ENH), 64 replicaspeak memory (GB): 92.5 -> 18.0Improved
apply FlashSchNet to a small proteinCGSchNet · Chignolinspeed (timestep*mol/s): 3578 -> 5222Improved
use FlashSchNet on the largest proteinCGSchNet · Alpha3Dspeed (timestep*mol/s): 288 -> 2610Improved
FlashSchNet optimizations versus CGSchNet on structural accuracyFlashSchNet · VillinGDT-TS: 0.78 -> 0.74No clear effect
FlashSchNet optimizations versus CGSchNet on native-contact fidelityFlashSchNet · Villinlargest metastable Q: 0.96 -> 0.88Hurt
use learned potential (FlashSchNet) instead of MARTINIMARTINI · VillinGDT-TS: 0.46 -> 0.74Improved
grow the neighbor graph during unfolding (denser edges) for CGSchNet's scatter aggregation (ablation)CGSchNet · 1ENH elongated 300k-step simulationthroughputHurt
use contention-free CSR segment reductions (ablation)FlashSchNet · 1ENH elongated 300k-step simulationthroughput stabilityImproved
reduce memory via IO-aware design to allow larger batch sizes (ablation)FlashSchNet · replica scaling, four proteinsmax replicas before OOM: 256 replicas (Alpha3D) to 2048 replicas (Chignolin)Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy