Brandon Wood's Recent LinkedIn Posts

UMA-S 1.2 is here! ~50% faster, ~40% more accurate on Open Molecules test set, and expanded data coverage for catalysts (oxides and interfaces), molecules, and polymers! We hope this release addresses a number of items on the collective wish list (definitely not all 😜). Details below. 𝗠𝗼𝗱𝗲𝗹 𝗵𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀: - 𝗕𝗲𝘁𝘁𝗲𝗿 𝗵𝗮𝗻𝗱𝗹𝗶𝗻𝗴 𝗼𝗳 𝗰𝗵𝗮𝗿𝗴𝗲, leading to improvements across molecular systems e.g. relative improvements to UMA-S-1.1 biomolecules ~50% and electrolytes ~30%.   - 𝗖𝗼𝘃𝗲𝗿𝗮𝗴𝗲 𝗼𝗳 𝗺𝗼𝗿𝗲 𝗮𝘁𝗼𝗺𝗶𝗰 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗮𝗻𝗱 𝗹𝗲𝘃𝗲𝗹𝘀 𝗼𝗳 𝘁𝗵𝗲𝗼𝗿𝘆. We added OC22 (oxide catalysts) and OC25 (electrolyte/inorganic interfaces). OMol25 data was expanded including polymers (OPoly26). Now ~520M total DFT calculations! - 𝗙𝗮𝘀𝘁𝗲𝗿 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝘀𝗽𝗲𝗲𝗱. ~50% speedup for UMA-S (with turbo mode) compared to our previous releases.  - 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀 𝗳𝗼𝗿 𝗸𝗻𝗼𝘄𝗻 𝗶𝘀𝘀𝘂𝗲𝘀: e.g. numerical and stability improvements for hessians and phonons, and more accurate diatomics and ionization potentials. 𝗟𝗮𝗿𝗴𝗲-𝘀𝗰𝗮𝗹𝗲 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗠𝗼𝗹𝗲𝗰𝘂𝗹𝗮𝗿 𝗗𝘆𝗻𝗮𝗺𝗶𝗰𝘀 (𝗠𝗗): - 𝗥𝘂𝗻 𝗹𝗮𝗿𝗴𝗲 𝘀𝗰𝗮𝗹𝗲 𝗠𝗗 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻𝘀 𝘄𝗶𝘁𝗵 𝗲𝗮𝘀𝗲: UMA is built to easily scale up to multi-node multi-gpu parallel inference with Ray (https://www.ray.io/) under the hood. Easily run MD with ASE and LAMMPs on large scale systems with ns/day speeds, battle tested to handle up to hundreds of GPUs and systems up to 1M atoms. UMA handles all the parallelism for you; no need to manually juggle dozens of packages and complex cpp builds with LAMMPs, Kokkos, MPI, CUDA etc. - 𝗥𝘂𝗻 𝗯𝗮𝘁𝗰𝗵𝗲𝗱 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻𝘀: A client–server Ray framework enables batched simulations—structural relaxations, molecular dynamics, etc—without requiring algorithm rewrites. We measured up to 4x speedups over serial calculations for batches of small systems on a single H100 GPU. 𝗦𝗵𝗼𝘂𝘁 𝗼𝘂𝘁 𝘁𝗼 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗙𝗔𝗜𝗥𝗰𝗵𝗲𝗺 𝘁𝗲𝗮𝗺 𝗳𝗼𝗿 𝗺𝗮𝗸𝗶𝗻𝗴 𝗶𝘁 𝗵𝗮𝗽𝗽𝗲𝗻!   Luis Barroso-Luque, Elisa Cascardi, Misko Dzamba, Ray G., Vahe Gharakhanyan, Daniel Levine, Kyle Michel,  Benjamin Miller,  Ammar Rizvi, Jagriti Sahoo, Muhammed Shuaibi, Zachary Ulissi, Larry Zitnick 𝗔𝗹𝘀𝗼 𝘁𝗵𝗲 𝘁𝗲𝗮𝗺 𝗶𝘀 𝗵𝗶𝗿𝗶𝗻𝗴 so if you’re excited about the work we are doing come and join us! https://lnkd.in/gdDc24nA Model: https://lnkd.in/gWww_kWU Code: https://lnkd.in/gfTqAXtS Paper: https://lnkd.in/gRnMuDru * Plot details: Sample multi-gpu inference speeds for UMA-S-1.2. ns/day assumes 1 fs time steps. The test system uses a fcc lattice with a=3.8A. Tests are run on H200 GPUs and include graph generation time.