Skip to content

Rh-Enamide

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

Data

Inputs: Rh-enamide training set (structures, Jaguar QM data, MM3 FF template)

Outputs: Benchmark results (JSON results for all backend × optimizer combinations)

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

flowchart LR
    A[Jaguar QM data] --> B[Seminario]
    B --> C[Initial FF]
    C --> D[MM Frequencies]
    D --> E[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: Minimize weighted sum-of-squares between QM and MM frequencies across all 9 molecules simultaneously

Functional forms

The MM3 force field template uses MM3 functional forms (cubic/quartic stretch, sextic bend), supported by OpenMM and Tinker. For JAX and JAX-MD engines, which only support harmonic potentials, we use a harmonic copy of the same Seminario-estimated parameters. Initial scores differ between functional forms because the energy expressions differ, but convergence behavior is comparable.

Results (2 iterations)

Results generated with q2mm-benchmark --system rh-enamide --max-iter 2. All results use the new BenchmarkResult JSON format.

JAX Engines (harmonic FF)

Backend Optimizer RMSD₀ RMSD Evals Time
JAX (harmonic) L-BFGS-B 18,177 36,105 1,648 50 s
JAX (harmonic) Nelder-Mead 18,177 82,597 186 8 s
JAX (harmonic) Powell — LinAlgError — —
JAX-MD (OPLSAA) L-BFGS-B 24,727 67,391 1,099 229 s
JAX-MD (OPLSAA) Nelder-Mead 24,727 68,857 187 39 s
JAX-MD (OPLSAA) Powell — LinAlgError — —

MM3 Engines (MM3 FF)

Backend Optimizer RMSD₀ RMSD Evals Time
OpenMM (CUDA) Nelder-Mead 19,342 78,134 187 172 s
OpenMM (CUDA) L-BFGS-B — too slow — >30 min
OpenMM (CUDA) Powell — too slow — >10 min

Key findings

  • OpenMM CUDA now works on Blackwell (RTX 5090) — install OpenMM-CUDA-12 pip package; the engine falls back gracefully to CPU if CUDA context creation fails
  • Nelder-Mead completes on all backends including OpenMM CUDA (172 s for 9 molecules × 182 params)
  • L-BFGS-B is impractical for OpenMM — frequency gradients use finite differences (183 evals per gradient step), making each L-BFGS-B iteration extremely slow
  • Powell crashes with LinAlgError on JAX/JAX-MD — line-search drives parameters to physically unreasonable values
  • All results use jac="auto" which auto-enables analytical gradients for engines that support them (currently energy-only; frequency gradients still use finite differences)

RMSD increases with only 2 iterations

With maxiter=2, Nelder-Mead does not have enough iterations to converge for 182-parameter systems. The RMSD increases because the simplex has barely started exploring. Use grad-simp cycling (below) for converged results.

Grad-Simp Cycling (converged)

Full optimization using L-BFGS-B (GRAD) → Nelder-Mead (SIMP) alternation with up to 5 parameters per cycle. Uses auto-generated harmonic FF from molecular topology (94 parameters across 9 molecules, 1,273 frequency reference values).

Device Cycles Evals Opt time Final score Improvement
GPU (RTX 5090) 3 30,637 1,117 s 34.56 98.40%
CPU 4 30,936 686 s 32.78 98.48%

Both devices achieve >98% improvement from the Seminario starting point with nearly identical eval counts (~30 k). CPU is 1.6× faster due to float64 overhead on consumer GPUs. See the GPU acceleration page for a detailed analysis of per-eval throughput, why CPU wins, and the path to making GPU viable.

Key result

grad-simp cycling reduces the score from 2,161 → ~33 (98.5% improvement) in 3–4 cycles. This confirms that the cycling optimizer and JAX backend can handle real organometallic systems end-to-end.

Data generated from Jaguar B3LYP/LACVP reference data in examples/rh-enamide/. Full results archived in benchmarks/rh-enamide/.