PES Scan with SD/CG Optimization

Overview

SD/CG is available for relaxed PES scans through #scan(method=sdcg). It reuses MAPLE's hybrid SD-to-CG optimizer at each scan point: the SD phase first stabilizes high-force geometries, then the CG phase improves convergence near the constrained minimum.

This method is useful when a scan contains both rough and smooth regions, or when you want a single optimizer that can handle changing scan-point difficulty.

Parameters

Per-point SD/CG settings reuse the same controls as #opt(method=sdcg). The scan driver keeps per-point optimizer verbosity quiet.

Parameter	Type	Default	Description
`max_step`	float	`0.2`	Maximum step length in Angstrom.
`max_iter`	int	`256`	Maximum optimization cycles per scan point.
`sd_max_iter`	int	`50`	Maximum iterations in the initial SD phase before switching to CG.
`cg_switch_fmax`	float	`0.0`	Switch from SD to CG when the maximum force drops below this value (Hartree/Angstrom). The default `0.0` enables the runtime auto-threshold: 0.5 × the initial maximum force.
`cg_restart_threshold`	float	`0.2`	Restart CG directions when the gradient change ratio exceeds this value.
`cg_beta_method`	str	`"prp+"`	Conjugate gradient beta formula.
`diis_enabled`	bool	`True`	Enable DIIS extrapolation acceleration.
`diis_store_every`	int	`5`	Store a DIIS snapshot every N steps.
`diis_min_snapshots`	int	`3`	Minimum snapshots required before DIIS extrapolation is attempted.
`diis_memory`	int	`6`	Maximum number of DIIS snapshots stored.
`verbose`	int	`0` in scans	Per-point optimizer verbosity. Scan keeps this low so the scan log stays readable.

Input Example

Checked-in MAPLE example: example/scan/sdcg/glycine_backbone_torsion.inp. This is the relaxed glycine backbone torsion scan shown in the visualization below:

#model=aimnet2nse
#scan(method=sdcg,mode=relaxed)
#device=gpu0

N     -1.64695360    -1.29691911    -0.90833498
C     -0.45216594    -0.59642464    -1.41625210
C     -0.84033589     0.54283507    -2.34787441
O     -1.97257790     0.87393427    -2.66478245
O      0.22929521     1.20487418    -2.83060947
H     -2.25086549    -0.60633338    -0.45686313
H     -2.19402852    -1.61340515    -1.71207045
H      0.17205835    -1.30686984    -1.96558531
H      0.10890908    -0.18794943    -0.57097223
H     -0.15752184     1.89889132    -3.40493371

S 6 1 2 3 5.0 72

The final S line is a dihedral scan: four atom indices followed by a 5.0 degree step and 72 increments, producing 73 scan points including the initial geometry.

Scan Visualization

The animation below pairs the generated scan geometries with the energy profile for a relaxed glycine backbone torsion scan using SD/CG.

Fig. 1 — Glycine backbone torsion relaxed scan (SD/CG)

When to Use SD/CG Scans

Scan points vary strongly in difficulty: The SD phase rapidly reduces large forces without requiring any history, then hands off to CG for efficient final convergence — combining the robustness of SD with the speed of CG.
Distorted scans where pure CG starts too aggressively: SD/CG is a robust default that avoids the early oscillations pure CG can show on rough geometries.
Well-behaved large scans: L-BFGS may still be faster for routine production grids when starting geometries are already reasonable.

Note

For production-quality scans, L-BFGS (the default) is generally preferred due to its superior per-point convergence near minima. SD/CG is most useful as a robust scan optimizer for severely distorted structures or when L-BFGS has difficulty converging at certain scan points.