iiinova — where your digital self keeps evolving

Rohit Goswami

@rgoswami.me

106 followers 609 following 8 posts

Computational chemist with a Bayesian taint. Working on foundational cross-language tools at @labcosmo.bsky.social‬. More @ rgoswami.me

Posts Replies Media

Rohit Goswami @rgoswami.me · 12/09/2026

Common data formats and model interfaces for atomistic machine learning Machine learning for molecular simulations has produced a large ecosystem of software, written in different languages (Python, C++, Fortran, Julia) and built on different frameworks (PyTorch, JAX, scikit-learn). doi...

010

Rohit Goswami @rgoswami.me · 12/09/2026

Geometry-aware data pruning halves wall time for GP-accelerated saddle searches Gaussian process (GP) regression can accelerate saddle point searches by building a surrogate energy surface on the fly, reducing the number of expensive quantum mechanical force evaluations. doi.org/10.1002...

000

Rohit Goswami @rgoswami.me · 12/09/2026

A map of the shapes a molecule visits, not just a graph of its energy On a cycloaddition, a cheap learned model and a slower standard calculation land on similar energy colours even though the atoms sit in different places. doi.org/10.1016/j.mex.2026.103851

000

Rohit Goswami @rgoswami.me · 12/09/2026

A reproducible NEB workflow, end to end Ad-hoc glue for endpoints and path setup, replaced by Snakemake: PET-MAD plus eOn, every dep from conda-forge. Validated on HCN to HNC. Same output on laptop and cluster. doi.org/10.1016/j.mex.2026.103899

000

Rohit Goswami @rgoswami.me · 12/09/2026

One Bayesian loop for minima, saddles, and reaction paths Minimization, dimer, and NEB are the same six-step surrogate loop. They differ only in the target and the acquisition rule. GP with gradients: ~10x fewer QM calls. doi.org/10.1021/acsphyschemau.6c000…

000

Rohit Goswami @rgoswami.me · 12/09/2026

A faster saddle point hunt that knows when to switch methods CI-NEB is reliable but expensive; MMF is cheap but drifts. Adaptive hybrid: median 57% fewer energy/force calls (Baker-Chan + PET-MAD), 31% on 59 Pt(111) heptamer transitions. doi.org/10.3389/fchem.2026.1807063

000

Rohit Goswami @rgoswami.me · 12/09/2026

rsx: Rust RADSex replacement. Same commands, RAM bound by sample count, Bayes grades per marker. 8.38x geometric-mean faster (56 pairs, 41.9e9 bases). CUDA up to 29.86x. doi.org/10.1186/s12859-026-06628-4

011

Reposted by Rohit Goswami

COSMO Lab @labcosmo.bsky.social · 22/08/2025

🚨 #machinelearning for #compchem goodies from our 🧑‍🚀 team incoming! After years of work it's time to share. Go check arxiv.org/abs/2508.15704 and/or metatensor.org to learn about #metatensor and #metatomic. What they are, what they do, why you should use them for all of your atomistic ML projects 🔍.

1148

Reposted by Rohit Goswami

COSMO Lab @labcosmo.bsky.social · 18/08/2025

Go metatensor.org!

A metatensor.org sticker on top of the mattehorn

071

Reposted by Rohit Goswami

COSMO Lab @labcosmo.bsky.social · 27/05/2025

📢 Running molecular dynamics with time steps up to 64fs for any atomistic system, from Al(110) to Ala2? Thanks to 🧑‍🚀 Filippo Bigi and Sanggyu Chong, with some help from Agustinus Kristiadis, this is not as crazy as it sounds. Let us briefly introduce FlashMD⚡ arxiv.org/html/2505.19...

Scheme of the GNN architecture of the FlashMD method.

13712