Production Saddle Search¶

This tutorial demonstrates GP-Dimer and OTGPD saddle point search on a real potential energy surface via RPC oracle.

Seeding from NEB¶

In practice, saddle point searches are rarely started from scratch. The common workflow is:

Run NEB to find the MEP between reactant and product
Identify the climbing image (highest energy image on the converged path)
Use the path tangent at the climbing image as the dimer orientation
Run GP-Dimer or OTGPD to refine the transition state geometry

The NEB tangent provides a much better initial orientation than a random guess, eliminating the rotation phase entirely (max_rot_iter = 0).

Running the example¶

cargo run --release --example hcn_dimer --features io,rgpot -- --method all

This runs both GP-Dimer and OTGPD on the system100 dataset using PET-MAD via RPC. The initial guess is the midpoint between reactant and product, with orientation along the reactant-to-product displacement vector.

GP-Dimer vs OTGPD¶

Both methods find the transition state, but differ in oracle strategy:

GP-Dimer: calls the oracle at every outer iteration (predictable cost, straightforward convergence)
OTGPD: calls the oracle only when the GP is uncertain (adaptive, can skip calls when the GP is confident)

For starting geometries near the saddle (the NEB refinement case), both methods perform similarly. For starting geometries far from the saddle (exploratory search), OTGPD’s adaptive threshold avoids wasting oracle calls on poor early GP predictions.

Configuration for molecular systems¶

let mut cfg = DimerConfig::default();
cfg.t_force_true = 0.1;        // force convergence (eV/A)
cfg.trust_radius = 0.05;       // conservative for molecules
cfg.max_rot_iter = 0;          // skip rotation (NEB tangent)
cfg.trust_metric = TrustMetric::Emd;
cfg.atom_types = atomic_numbers.clone();
cfg.const_sigma2 = 1.0;        // constant kernel

Key differences from the LEPS example:

Smaller trust_radius(0.05 vs 0.1): molecular surfaces have finer features near the saddle
EMD trust metric: appropriate for molecular systems
const_sigma2= 1.0: stabilizes predictions for many-atom systems
No rotation: provided orientation from NEB tangent

When to use each method¶

Starting geometry	Method	Reason
NEB climbing image	GP-Dimer	Simple, predictable, few calls
Approximate TS	OTGPD	Adaptive threshold saves calls
Random guess	OTGPD	Better early-phase behavior
Known eigenmode	GP-Dimer	Direct with `max_rot_iter=0`