Molecular dynamics without hand-written input files: ask for a material property in chat, and LAMMPS with UMA runs it and returns the error bar

A molecular dynamics calculation of a material property can now be completed inside a conversation: describe the quantity you want, and the calculation is set up, run and returned with its error bar, units and qualifying warnings. The same calculation previously required installing software, writing an input file and submitting a queue script. The service runs on nanoHUB, with Claude, ChatGPT or the nanoHUB chat connected to the server executing LAMMPS and UMA on the user's behalf, so the numbers come from real calculations rather than description.

The assistant can read warnings to judge whether a run is trustworthy, rerun it larger or longer when a cell was too small or a trajectory too short, and compare the two results. That iteration happens without the user touching a file. The service offers compare_runs to place several runs side by side, and skills such as md-sanity covering equilibration, error bars, drift and finite size.

Every result names the potential behind it and the reference it should be compared with: a classical potential fitted to experiment reproduces experiment, and a machine-learned one fitted to PBE reproduces PBE. A simulation reports a property of the potential rather than of the material, and the two are often confused. For the equation of state of silicon computed with Stillinger-Weber, SNAP and UMA, the spread between them is the lesson.

The service is intended for students meeting molecular dynamics for the first time, who get a working calculation and an honest reading of it; for instructors who want a homework problem that can be explored; and for researchers who want a quick, bounded answer from a real code. Calculations run inside the user's own nanoHUB session, confined to their own directory, and each is reproducible from the input file it generated. Short work starts immediately; a long research run is not refused but its cost is put to the user first; work too long for any session goes to a cluster as one job after asking.

The material stresses that a simulation reports a property of the potential, not of the material, so results must be compared against the reference the potential was fitted to. If no potential is named, a classical one is chosen and a warning says so; machine-learned potentials are never chosen silently because they are slow. SNAP costs about a hundred times EAM.