Benchmarks

Reading guide

If you're new to q2mm, start with the Tutorial — this section is evidence and analysis, not a learning path. For per-system results and literature context, see the Systems section.

Benchmarks help distinguish software-path coverage from scientific fit quality and historical timing measurements. q2mm's benchmark program answers three questions:

1. Which publication problems can be built? What does each objective include or omit? → Publication Reproduction Coverage
2. Does the optimizer work? Given QM reference data and a starting FF, can q2mm reduce the penalty function? → Optimizer Comparison, Small Molecules
3. Is it fast enough? When does GPU or the analytical-gradient path pay for itself? → GPU Acceleration, When Analytical Wins

Benchmark sections

Section	What it covers
Publication Reproduction Coverage	Current profile/start records, target completeness, and source-linked limitations; not numerical validation
QFUERZA Validation	Implementation comparisons against the paper and Zenodo data, and starting-point diagnostics
Optimizer Comparison	Historical mixed-source optimization measurements with endpoint provenance and local-basin limits
GPU Acceleration	Historical CPU-vs-GPU wall-clock measurements and their artifacts
When Analytical Wins	Historical analytical-vs-finite-difference timing measurements and their artifacts

Registered systems

This inventory is generated from registered-system metadata. Registration does not imply that external inputs are installed or that a publication objective is complete. Follow the profile/start records rather than treating a system as one validated/not-validated label.

Registered system	Default functional forms	Publication coverage
CH3F (ch3f)	harmonic, mm3	No canonical publication record
F⁻ + CH3F SN2 TS (ch3f-sn2)	harmonic, mm3	No canonical publication record
Rh-enamide (rh-enamide)	mm3	Profile/start records
Heck relay (heck-relay)	mm3	Profile/start records
Pd-allyl amination (pd-allyl)	mm3	Profile/start records
Pd 1,4-conjugate addition (pd-conjugate)	mm3	Profile/start records
Rh 1,4-conjugate addition (rh-conjugate)	mm3	Profile/start records
Ferrocene ground-state force field (ferrocene)	mm3	Profile/start records

What is not covered yet

• Type-normalized penalties — the upstream Q2MM divides each data type by its count (score / N_type). Our un-normalized eigenmatrix loss creates a flat landscape that limits L-BFGS-B convergence.
• Selectivity prediction — q2mm does not yet have the conformational-search infrastructure needed for end-to-end enantioselectivity benchmarks.

Metrics used in this section

Metric	Meaning
R²	Coefficient of determination on eigenvalue scatter — 1.0 = perfect, 0 = no correlation, negative = worse than the mean.
RMSD	Root-mean-square deviation between QM and MM frequencies (cm⁻¹). Lower is better.
TS	Transition state — the highest-energy geometry along a reaction pathway.
FF	Force field — the set of parameters defining how atoms interact.
TSFF	Transition state force field — parameterized specifically for TS geometries.

Artifacts

Published-force-field re-evaluation fixtures are archived in test/fixtures/published_ff/. Historical measurements are separate from the current inventory: the linked analysis pages identify their tracked artifacts in q2mm-data.