GROMACS Guru with Memory-AI GROMACS simulation assistant

Example

A user runs `gmx mdrun` and gets a fatal error referring to inconsistent atom counts or a constraint solver failure. The assistant provides a stepwise diagnostic checklist: (1) compare atom counts in the coordinate file and the topology ([ atoms ] count), (2) inspect include files and .itp includes for mismatched residues or missing directives, (3) check for `constraints` versus `consref` mismatches and inappropriate LINCS settings, (4) re-generate the topology for the problematic molecule (e.g., via `pdb2gmx` or the chosen force-field tool) and re-check. For energy-drift issues the assistant suggests isolating the cause (short NVE test with tightly constrained bonds, lower dt, check thermostat/barostat groups) and explains which log lines and output to inspect (energies, pressure, temperature fluctuations).

Scenario

A postdoc preparing a ligand–protein production run observes a sudden crash reported as an index or atom mismatch. The assistant guides them to open the .top and .gro files, count atoms in the [ atoms ] block and the coordinate file (text editor or small awk/python snippet), identifies that a residue insertion step removed a terminal hydrogen, and shows how to regenerate the topology for that residue and re-run a short minimization to confirm the fix.

Example

The assistant stores and reuses exact commands and selection choices supplied by the user. For example, if you provided and preferred the RDF call: `gmx rdf -f traj.xtc -s topol.tpr -o rdf.xvg -seltype whole_mol_com -selrpos whole_mol_com`, the assistant will recall that selection and provide ready-to-run loops for batch analysis, e.g.: `for f in *.xtc; do gmx rdf -f "$f" -s topol.tpr -o rdf_${f%.xtc}.xvg -seltype whole_mol_com -selrpos whole_mol_com; done`. It also converts remembered commands into portable script fragments (Bash, Makefile, or simple Python wrappers) and annotates them with the assumptions made (units, index groups, expected topology).

Scenario

A user analyzed one trajectory interactively and later needs to process 120 replicate trajectories produced by a replica-exchange or ensemble campaign. Because the assistant remembered the chosen selection flags and index group names from the interactive run, it auto-populates the batch script and suggests a small verification step (run on the first 3 trajectories and compare results) to ensure consistent behavior across replicates.

Example

The assistant provides concrete, literature-backed recommendations for mdp settings and parallelization strategies: (a) recommend `dt=0.002` ps (2 fs) when bonds to hydrogens are constrained, otherwise use 1 fs; (b) advise LINCS parameters (e.g., `lincs_order` and `lincs_iter`) appropriate for complex restraint geometries; (c) recommend PME grid spacing ~0.11–0.14 nm and cutoffs near 1.0–1.2 nm depending on force field; (d) explain the trade-off between neighbor-list frequency (nstlist) and Verlet/skin settings, and to run short performance test runs to measure ns/day under different MPI/OpenMP/GPU decompositions. The assistant provides small mdp examples for common stages (em, nvt, npt) and highlights which parameters to change for special cases (membrane, coarse-grained, polarizable force fields).

Scenario

A compute cluster user preparing a 200k-atom membrane protein simulation wants maximum throughput on a GPU node. The assistant recommends running a short (50–100 ps) benchmarking experiment with two configurations: (1) more MPI ranks with fewer threads, (2) fewer MPI ranks with more OpenMP threads, and suggests which metrics to record (wallclock time, ns/day, energy drift). It also explains how to choose PME grid spacing and why FFT grid factors matter, and it provides an mdp snippet for stable production runs (integrator, dt, constraints, thermostat/barostat, PME settings) along with a checklist for reproducibility (fixed random seeds, documented versions, deffnm naming conventions).