Usage: mlmm opt [OPTIONS]
ML/MM geometry optimization with L-BFGS or RFO.
Options:
-v, --verbose LEVEL Console verbosity 0-3 (default 2). 0=silent;
1=milestones only; 2=+optimizer cycle tables,
per-stage timing, VRAM, deliverable paths;
3=everything (full config blocks, per-file
paths, DEBUG logging). [0<=x<=3]
--help-advanced Show all options (including advanced settings)
and exit.
-i, --input FILE Input structure file (PDB/mmCIF, or XYZ). XYZ
provides higher coordinate precision. If XYZ,
use --ref-pdb to specify PDB topology for atom
ordering and output conversion. [required]
--ref-pdb FILE Reference PDB topology when input is XYZ. XYZ
coordinates are used (higher precision) while
PDB provides atom ordering and residue
information for output conversion.
--parm7, --parm FILE Amber parm7 topology covering the whole enzyme
complex. [required]
--model-pdb FILE ML-only, link-H-free PDB subset; atom
identity/order must match the full PDB/parm7.
When provided, it defines ML membership;
--detect-layer still reads valid
movable/frozen MM B-factors.
--model-indices TEXT Comma-separated atom indices for the ML region
(ranges allowed like 1-5). Used when --model-
pdb is omitted.
--freeze-atoms TEXT Comma-separated 1-based atom indices to freeze
(e.g., '1,3,5').
--hessian-cutoff, --hess-cutoff FLOAT
Distance cutoff (Å) from ML region for MM
atoms to include in Hessian calculation.
Applied to movable MM atoms and can be
combined with --detect-layer. [default: (all
movable MM atoms)]
--movable-cutoff FLOAT Distance cutoff (Å) from ML region for movable
MM atoms. MM atoms beyond this are frozen.
Providing --movable-cutoff disables --detect-
layer and uses distance-based layer
assignment. [default: (use freeze_atoms)]
--distance-restraint, --dist-freeze TEXT
Distance restraints: inline Python literal
(e.g. '[(1,5,1.4)]') or a YAML/JSON spec file
path. Format: (i,j,target_Å) triples. Target
may be omitted to freeze at the current
distance: (i,j).
--one-based / --zero-based Interpret --distance-restraint indices as
1-based or 0-based. [default: one-based]
--restraint-k, --bias-k FLOAT Harmonic restraint strength k [eV/Å^2] for
--distance-restraint. [default: (300.0)]
--max-cycles INTEGER RANGE Maximum number of optimization cycles.
[default: (100000); x>=1]
--dump / --no-dump Write optimization trajectories
('optimization_trj.xyz' and
'optimization_all_trj.xyz'). [default: no-
dump]
-o, --out-dir TEXT Output directory. [default: ./result_opt/]
--thresh [gau_loose|gau|gau_tight|gau_vtight|baker|never]
Convergence preset. [default: (gau)]
--opt-mode [grad|hess|lbfgs|rfo]
Optimization mode: grad/lbfgs or hess/rfo.
[default: grad]
--microiter / --no-microiter Alternate one RFO step with MM L-BFGS
relaxation. Used in hess mode without
embedding. [default: microiter]
--flatten / --no-flatten Enable/disable imaginary-mode flatten loop
after optimization. [default: no-flatten]
--reject-uphill / --no-reject-uphill
Opt in to rejecting uphill RFO trials in hess
mode (tolerance: 1e-4 Hartree) and final-check
the retained geometry at the emergency floor.
Ignored in grad/lbfgs mode. [default: no-
reject-uphill]
--config FILE Base YAML configuration file applied before
explicit CLI options.
--show-config / --no-show-config
Print the loaded YAML file and its top-level
keys, then continue. [default: no-show-
config]
--dry-run / --no-dry-run Validate options and inputs without running
optimization. [default: no-dry-run]
--convert-files / --no-convert-files
Convert XYZ/TRJ outputs into PDB companions
based on the input format. [default: convert-
files]
-b, --backend [uma|orb|mace|aimnet2|dft]
High-level backend for the ONIOM model region.
[default: (uma)]
--embedcharge / --no-embedcharge
Enable electrostatic embedding. MLIP backends
use the computationally expensive xTB point-
charge delta correction; dft uses native PySCF
MM point charges. [default: no-embedcharge]
--embedcharge-cutoff FLOAT Distance cutoff (Å) from the ML region for MM
point charges used by embedding. [default:
(12.0)]
--link-atom-method [scaled|fixed]
Link-atom position mode: scaled (g-factor) or
fixed (1.09/1.01 Å). [default: (scaled)]
--mm-backend [hessian_ff|openmm]
MM backend. MM Hessians use finite differences
by default; set calc.mm_fd: false for the
hessian_ff analytical path. [default:
(hessian_ff)]
--mm-only / --no-mm-only Skip the MLIP component entirely and minimize
using only the MM force field on the full
system. Layers (movable/frozen) are still
honored via B-factor encoding or --movable-
cutoff. Only --opt-mode grad (L-BFGS) is
supported in this mode; microiteration is
automatically disabled. [default: no-mm-only]
--cmap / --no-cmap Preserve CMAP terms in both real and model MM
layers when present in parm7. [default:
(cmap)]
--out-json / --no-out-json Write machine-readable result.json to out_dir.
[default: no-out-json]
--detect-layer / --no-detect-layer
Automatically detect ML/MM layers from input
PDB B-factors (ML=0, MovableMM=10,
FrozenMM=20) when explicit ML membership is
absent. With explicit membership, retain valid
movable/frozen MM B-factor layers. [default:
detect-layer]
-q, --charge INTEGER ML region charge. Required unless --ligand-
charge or YAML calc.model_charge gives it.
-l, --ligand-charge TEXT Total charge for unknown ligand residues or a
per-resname mapping (e.g., GPP:-3,SAM:1), used
to derive the ML-region charge when -q is
omitted (requires PDB/mmCIF input or --ref-
pdb).
-m, --multiplicity INTEGER RANGE
Spin multiplicity (2S+1) for the ML region.
[default: (1); x>=1]
--coord-type [cart|redund|dlc|tric]
Optimization coordinate system
(cart|redund|dlc|tric). cart is the default;
command-specific choices are listed here.
[default: (cart)]
--print-every INTEGER RANGE Print optimizer status every N cycles.
[default: (100); x>=1]
--precision [fp32|fp64] MLIP backend precision: fp32 or fp64. Unset
defaults per backend (uma: fp32; orb, mace:
fp64). Routed to backend-specific kwargs (UMA
precision / ORB precision / MACE
default_dtype). aimnet2: fp32 no-op; fp64
rejected. [default: (per backend: uma fp32;
orb, mace fp64)]
--uma-workers, --workers INTEGER
MLIP predictor workers (UMA). >1 uses a
parallel predictor (fairchem-core[extras]);
combining it with an analytical Hessian is an
error. Default 1. [default: (1)]
--uma-workers-per-node, --workers-per-node INTEGER
Workers per node when the parallel MLIP
predictor is used (--workers > 1). [default:
(1)]
--backend-model TEXT Model variant for the selected --backend (e.g.
uma-s-1p2 / uma-m-1p1 for uma,
orb_v3_conservative_omol for orb, MACE-OMOL-0
/ off:small for mace). [default: (the
selected backend's own model)]
--calc-file FILE Python file exposing get_calculator(...) -> an
ASE Calculator used as the ML-region backend
(overrides --backend). Couples GFN-xTB / DFTB+
/ any ASE engine. See --calc-file-func-name.
--calc-factory, --calc-file-func-name TEXT
Name of the callable in --calc-file that
returns an ASE Calculator (or a module-level
Calculator instance). CLI overrides config
YAML; otherwise defaults to get_calculator.
[default: (get_calculator)]
--deterministic / --no-deterministic
Request deterministic algorithms for
controlled operations; verify exact
reproducibility on the complete target stack.
[default: no-deterministic]
--allow-charge-mult-mismatch Skip the ML-region charge/multiplicity
electron-parity check (logs that it was
skipped). An open-shell ML region needs a
matching multiplicity; use this only for an
intentional nonstandard input such as a
covalently-cut region.
--func-basis TEXT High-level method as FUNCTIONAL/BASIS;
HF/BASIS is accepted. [default:
(wb97m-v/def2-svp)]
--dft-engine, --engine [gpu|cpu]
PySCF execution engine used by --backend dft.
[default: (gpu)]
--save-scf-checkpoint / --no-save-scf-checkpoint
Persist a structure-bound PySCF checkpoint.
[default: (disabled)]
--scf-checkpoint FILE Load/save the optional structure-bound PySCF
checkpoint at PATH.
--dft-low-memory, --lowmem / --no-dft-low-memory, --no-lowmem
Use GPU4PySCF rks_lowmem for closed-shell GPU
DFT; open-shell GPU and CPU use standard
direct JK. --no-lowmem enables density
fitting. [default: (lowmem)]
--scf-stepwise-grid / --no-scf-stepwise-grid
Converge the first SCF on a coarse grid, then
on the final grid (later SCFs reuse the
previous density as usual). [default:
(enabled)]
--dft-nprocs INTEGER RANGE PySCF/OpenMP CPU threads; GPU count is
unaffected. [default: (auto); x>=1]
--dft-memory, --dft-mem TEXT PySCF host RAM limit (for example 64GB or
120000MB). [default: (auto)]
--stop-plateau / --no-stop-plateau
Stop when the energy stops changing while the
convergence criteria are still unmet, and
report the run as stalled. It never signals
convergence; an explicit --max-cycles remains
the hard bound. The MM micro iterations are
never stopped this way. [default: no-stop-
plateau]
--stop-plateau-thresh FLOAT Energy range (hartree) below which --stop-
plateau treats the window as flat. [default:
(1e-4)]
--stop-plateau-window INTEGER Number of consecutive cycles --stop-plateau
inspects. [default: (50)]
-h, --help Show this message and exit.
