Trinity science campaigns¶
These are real campaigns run end-to-end through Trinity's chat interface — DFT, molecular dynamics, ML benchmarking, and infrastructure verification — each backed by a full execution trace and, for auditability, a plain-language scientist report.
-
N2 gas molecular dynamics — two models, one open-ended ask
One vague prompt, two real experiments: a clean ideal-gas sweep and a two-site model that overshot temperature.
-
Fixing OpenFold's GPU kernel
A silently CPU-only fused-attention kernel diagnosed, rebuilt for CUDA, and verified against a reference structure.
-
MLPerf Storage vector-database benchmark: DAOS vs. Lustre
An overnight Aurora benchmark comparing storage backends and scaling query throughput across client nodes.
-
Closing out Polaris application verification
Three stubborn code failures — NWChem, QMCPACK, WRF — diagnosed and fixed to complete a full verification suite.
-
First job on Perlmutter
A cross-facility smoke test that surfaced two real infrastructure bugs before declaring the system verified.
-
3C-SiC ground-state energy via Quantum ESPRESSO
A clean first-try DFT run that still flagged its own smearing choice and residual-pressure caveat.
-
Liquid argon MD — catching its own mistake
A "liquid argon" run that was actually supercritical fluid, caught and corrected by Trinity itself.
-
Graphene band structure — a 3-round convergence study
Three rounds of DFT refinement resolving k-point convergence and confirming graphene's Dirac cone.
-
Predicting Ti K-edge XANES spectra across three TiO2 polymorphs
ML-predicted spectra for anatase, brookite, and rutile, correctly matched against experimental standards.