Skip to content

Rh 1,4-Conjugate

Rh 1,4-conjugate addition is the thesis-era Rh analogue of the Pd composed-force-field systems: a base MM3 field plus OPT overlay that does not transfer cleanly under our engine.

Scope

  • Type: Transition state (Rh-catalyzed 1,4-conjugate addition)
  • Molecules: 10 TS structures
  • Parameters: 488 (OPT substructure: 24 bonds, 46 angles, 348 torsions)
  • QM reference: B3LYP-D3/6-31G(d)

Publication

Property Value
Thesis Wahlers, J. Ph.D. Dissertation, University of Notre Dame, 2022, Ch. 6
DOI —
System Rh-catalyzed 1,4-conjugate addition
Training set 10 transition-state structures
Engine MacroModel MM3*

What the thesis reports

What the original Q2MM workflow fitted

The Chapter 6 Rh systems continue the same Q2MM strategy: a MacroModel MM3* transition-state force field fit against multiple data types, not just eigenvalues.1

  • Structural targets
  • Hessian/eigenvalue targets
  • MacroModel MM3* optimization
  • External selectivity validation on literature examples

Reported outcomes

Wahlers reports separate internal-fit ranges for two ligand classes:1

  • Bisphosphine systems: slopes 0.94–1.01, R² 0.91–0.99
  • Diene systems: slopes 1.0–1.07, R² 0.92–0.99
  • Bisphosphine selectivity validation: MUE 4.1 kJ/mol, R² = 0.64, 67 structures
  • Diene selectivity validation: MUE 5.3 kJ/mol, R² = 0.37, 69 structures

Our reproduction

Metric Value
Overall eigenvalue R² -3.90
Per-molecule R² range all negative
Positive R² values 0 / 10
Aggregate frequency RMSD 645.7 cm⁻¹ (per-molecule avg: 228.3)

What this means: A negative R² means our engine's reproduction of the published eigenspectrum is worse than simply predicting the average — a complete failure of cross-engine transfer, not a small miss.

Negative across the full training set

All per-molecule R² values are negative. Under our engine, the reproduced eigenspectrum is not preserving the literature fit at all.

Benchmark results

Ratio gate now passes — loader API refactor + MM3 angle gradient fix

The pre-refactor loader silently overwrote the Wahlers OPT parameters with raw QFUERZA projections, sending JaxLoss's inner geometry minimization into pathological regions and producing ratios that varied wildly across runs (0.46 / 0.96 / ~4 × 10³ in successive sessions). After the loader API refactor that preserves the published OPT values as-is, the ratio is 1.00 — comfortably inside the [0.85, 1.15] band. JaxLoss-guided optimization is now possible.

Run with --n-evals 10 so the verdict is statistically defensible against the per-call engine noise documented in #284. Numbers below are from the post-fix (q2mm MM3 angle gradient-correctness patch) run.

Metric Value
Ratio check 0.996 (in_band)
Initial ObjectiveFunction (n=10 mean) 6.293 × 10⁶ ± 2.60 % CI₉₅
Final ObjectiveFunction (n=10 mean) 5.160 × 10⁶ ± 1.92 % CI₉₅
Improvement (mean Δ%) 18.00 % (SIGNIFICANT — CI₉₅ ± 4.17 %)
L-BFGS-B iterations / OF evaluations 2 / 2
Optimizer L-BFGS-B (scipy) over JaxLoss analytical gradients
Wall time 691 s opt + ~13 min for 20 post-eval samples

Per-category fit before and after optimization (single GPU calls; per-category R² varies by ~1–2 % across calls — see the noise context below):

Category n_refs R² (optimized)
bond_length 457 0.822
bond_angle 926 0.540
eig_diagonal 1,244 −12.85

Newly unlocked: 18 % real-OF reduction after MM3 angle gradient fix

A previous baseline (q2mm-data#9) reported −0.080 % ± 1.18 % CI₉₅ for this system — "NOT SIGNIFICANT, FF sits at a JaxLoss local minimum". That conclusion was wrong. The optimizer wasn't finding a real local minimum; it was hitting a spurious stationary point caused by a gradient correctness bug in the JAX angle term (arccos(clip(cos θ)) killed gradient information at near-collinear geometries — see #284 for the diagnosis). After the fix replaces the clip with a custom-VJP atan2-based angle function, L-BFGS-B finds a real descent direction worth 18.00 % ± 4.17 % CI₉₅ in the real ObjectiveFunction (and the surrogate score itself drops by 18 % in tandem, confirming the fix made JaxLoss honest about the actual descent direction). The 4602-ratio non-determinism reported in an earlier session is also resolved (ratio now stable at 1.00 across runs; closes #278).

The eigenmatrix R² remains negative — the optimizer improved the real ObjectiveFunction substantially but the cross-engine MM3* ↔ JAX-engine eigenmatrix gap that affects all Wahlers systems is a separate, deeper issue (see "Comparison and gap analysis" below). bond_length and bond_angle R² went slightly down from the pre-optimization values (0.89 → 0.82, 0.45 → 0.54) because the optimizer traded a fraction of geometry-target fit for the much larger eigenmatrix improvement that drove the overall ObjectiveFunction down by 18 %.

The dramatic improvement vs the pre-refactor per-category numbers (where bond_length R² was −58) is the loss of the QFUERZA overwrite that was destroying the published Wahlers fit. The eigenmatrix R² is still negative, reflecting the same MM3* ↔ JAX-engine cross-engine gap that affects all Wahlers systems.

See Optimizer Comparison for the cross-system comparison. Raw numbers (published-start baseline) are in the from-published/ baseline in ericchansen/q2mm-data, with full provenance (q2mm git SHA, JAX/OpenMM device, ratio_tol, timestamp). Canonical QFUERZA-start results (default since q2mm#290) live at convergence/ and are summarized in the QFUERZA-recovery doc.

Comparison and gap analysis

Comparison

The thesis reports respectable-to-strong internal fits across both ligand classes. Our reproduction does not transfer that quality.

As with the Pd systems, this is a composed-force-field transfer problem. The Rh 1,4-conjugate TSFF combines a base MM3 field with an OPT overlay, and that composition is sensitive to engine-specific semantics.

The optimizer story is mixed but still informative: L-BFGS achieves 331.3 cm⁻¹, and Optax Adam reaches 307.0 cm⁻¹, yet the reproduced eigenspectrum remains negative across the entire training set. Better optimizer robustness helps the benchmark objective; it does not remove the transfer gap.

Gap analysis

To close the gap for Rh 1,4-conjugate addition, we would need:

  1. A verified composition path for the base MM3 field plus OPT overlay.
  2. Closer parity for Rh-specific MM3* behavior at the metal center.
  3. A re-fit against the original multi-target Q2MM objective only after the composed starting field behaves as intended.

The negative R² reflects a real transfer gap in the composed FF workflow.

Reproduce

Configure both Q2MM_SUPPORTING_INFO and Q2MM_MM3_BASE as described in External data for published systems before running this command.

python -m q2mm.diagnostics.cli --system rh-conjugate --backend jax --optimizer optax-adam

Raw data: q2mm-data/benchmarks/rh-1,4-conjugate-addition/.

  1. Wahlers, J. Ph.D. Dissertation, University of Notre Dame, 2022, Ch. 6. The chapter-level ranges are also summarized in Published FF Validation. ↩↩