Three published papers.
One thesis.
From supercritical fluids in chemical physics to 3D self-supervised vision transformers at Harvard. Each paper, the journal, and a link.
SpatialDINO: Self-Supervised Learning for 3D Vision Transformers
Arkash Jain et al. (Kirchhausen Lab, Harvard Medical School)
A 3D self-supervised vision transformer that beats a Nobel laureate-led approach for understanding subcellular structures from cryo-electron tomograms.
Read paper →Close-up of Vesicular ER Exit Sites by Volume Electron Imaging using FIB-SEM
Kirchhausen Lab
Volumetric reconstruction of mammalian ER exit sites at unprecedented resolution via FIB-SEM and learned segmentation.
Read paper →Ultrafast 2DIR comparison of rotational energy transfer, isolated binary collision breakdown, and near critical fluctuations in Xe and SF6 solutions
Arkash Jain et al.
First-author work on supercritical-fluid dynamics using ultrafast two-dimensional infrared spectroscopy.
Read paper →The full ML stack
What it takes to train SpatialDINO end-to-end.
Training stack
- › Infiniband / RDMA collective ops
- › RAID storage tier with NVMe cache
- › NVLink intra-node, DGX A100/H100 nodes
- › PyTorch FSDP + bf16 mixed precision
- › Activation checkpointing for large models
Open source
PyTorch Issue #144779
Diagnosed and reported a Rendezvous (RDZV) backend issue affecting Infiniband multi-node training; contributed reproduction steps and root-cause analysis.
View issue →