First Ascent by
Vayle Vera Cruz
The UMA Benchmark Ridge
Computational Chemistry / ML Force Fields
The Proposed Route
A high-exposure ridgeline comparing two views of the same mountain: the slow, rigorous DFT path (ORCA) and the blazingly fast ML interatomic potential path (Meta's UMA/eSEN). The climber benchmarks Gibbs free energies and transition state barriers, asking whether the new-school route gets you to the same summit as the old-school one.
๐ง The Crux
This is explicitly a benchmarking route, not a model-training route โ and that's a legitimate and important mode of ML research. The crux is rigor: absolute errors must be computed against both DFT and experimental (NIST) references. DFT calculations are the slow pitch โ even with B3LYP/def2-SVP optimization + single-point refinement, timelines are real.
โ ๏ธ Pre-Climb Checklist
โ UMA model accessible via fairchem GitHub + ASE. โ ORCA available on cluster. โ ๏ธ Explicitly frame this as benchmarking in your notebook intro โ state clearly that the ML model (UMA) is the subject of evaluation, not something being trained. โ ๏ธ Specify the molecule set before starting โ number and identity of test molecules should be locked in early.
Guidance
- Asks a question the field cares about: can we trust ML potentials for thermochemistry?
- Intellectual honesty is key โ report agreement AND disagreement with caveats
- Disagreement is equally interesting and publishable
- Centerpiece deliverable: results table (molecule | DFT ฮG | UMA ฮG | Exp ฮG | Error)
- That table is the summit photo
Source proposal: vayle_vera_cruz_final_proposal_chem269.pdf
CHEM 169/269 ยท Applied AI & Machine Learning for Biochemistry