Skip to content

Rh-Enamide — Jaguar Reference Data

Full-loop validation on a real organometallic system using Jaguar B3LYP/LACVP** QM reference data.

System Description

The rh-enamide training set consists of 9 transition-state structures for a Rh(I)-diphosphine catalyzed enamide hydrogenation. Each structure has 36 atoms including Rh, P, N, O, C, and H — a challenging test for the Seminario method and MM force field optimization.

Property Value
Structures 9 TS geometries
Atoms per structure 36
Elements Rh, P, N, O, C, H
QM level B3LYP/LACVP** (Hay-Wadt ECP for Rh)
QM program Jaguar (Schrödinger)
FF template MM3 (mm3.fld with Rh parameters)
Parameters 182 (8 bond, 23 angle, 36 vdW types)

Pipeline

graph LR
    A[9 Jaguar .in files] --> B[Seminario Estimation]
    B --> C[Initial Force Field<br/>182 parameters]
    C --> D[OpenMM Frequency<br/>Objective Function]
    D --> E[Nelder-Mead Optimizer]
    E --> F[Optimized FF]
  1. Load: 9 structures from MacroModel .mmo + Jaguar Hessians
  2. Seminario: Estimate bond/angle force constants from QM Hessians using the MM3 template (preserves vdW parameters for all atom types including Rh)
  3. Reference: Build multi-molecule frequency reference data — each molecule contributes its real vibrational frequencies (>50 cm⁻¹)
  4. Optimize: Nelder-Mead minimizes the weighted sum-of-squares between QM and MM frequencies across all 9 molecules simultaneously

Results

Seminario Estimation

Metric Value
Parameters 182 (8 bond, 23 angle, 36 vdW)
Time 0.033 s
Negative FCs C-H bond (5-17) in all 9 structures (TS reaction coordinate)

Optimization

Metric Value
Method Nelder-Mead (500 max iterations)
Frequency references 1,030 across 9 molecules
Initial score 434,172
Final score 101,077
Improvement 76.7%
Wall time 369 s (~6 min)

L-BFGS-B diverges on this problem

L-BFGS-B with finite-difference gradients (2×182+1 = 365 evaluations per gradient step) actually worsened the score from 434k to 1.5M in 3 iterations. The objective landscape for 182 parameters with 1,030 frequency references is too rough for finite-difference gradients. Nelder-Mead converges reliably without gradients.

Data generated by test/integration/test_full_loop_parity.py.