Publication evidence

Recorded on 2026-07-24 with the release build. This document supports an educational simulation-optimization tutorial. It does not validate the model for ventilation engineering.

Experimental design

Every objective evaluation solves the flow once and reuses it for all three training releases. Each pollutant objective is the worst of those releases. The three held-out releases are evaluated only after the optimizer selects a representative.

MODE and MAP-Elites both use 157 complete batches of 128, giving equal search budgets. MODE receives four objectives and four constraints. MAP-Elites minimizes the reporting scalar inside each flow/low-velocity niche and rejects infeasible candidates.

Three-seed results

Summary values use sample standard deviation:

The fixed baseline has training quality 1.598712 and held-out quality 1.791533. Relative to that baseline, mean held-out reporting quality improves by approximately 16.7% for MODE and 14.4% for the best MAP-Elites elites. These are descriptive results from three seeds, not confidence intervals.

MODE is more stable under this reporting scalar. MAP-Elites has a different purpose: its archive retains hundreds of distinct flow behaviors rather than only driving one scalar or preserving an objective Pareto set.

Search time covers the parallel optimization loop. Final population/archive reproduction, three held-out evaluations, CSV output, and plotting are outside that timer. Throughput is machine-specific and is not presented as a cross-machine benchmark.

The machine-readable source is replication-summary.csv.

Seed-42 representatives

MODE

training quality       = 1.120668469
held-out quality       = 1.491459638
training exposure      = 0.566551495
training max receptor  = 0.505681717
fan-power proxy        = 0.921422963
training final mass    = 0.233983046
held-out final mass    = 0.532366349
flow rate              = 1.932319 m²/s
low-velocity fraction  = 0.057604
flux mismatch          = 0.010783
flow residual          = 2.951909e-4

design = [
  0.2624119813644679,
  0.4348990083590912,
  0.2633892871687757,
  0.3914347733372533,
  1.4810477598402265,
  0.6099844574061106,
  0.6394449558711917,
  0.6449705049924965,
 -0.8515328431697532
]

MAP-Elites

training quality       = 1.143124392
held-out quality       = 1.490527843
training exposure      = 0.570405041
training max receptor  = 0.507570305
fan-power proxy        = 0.990489130
training final mass    = 0.241672745
held-out final mass    = 0.498903302
flow rate              = 2.025000 m²/s
low-velocity fraction  = 0.027586
flux mismatch          = 0.019638
flow residual          = 4.144476e-4

design = [
  0.2967337660939455,
  0.4500000000000000,
  0.3664460873093897,
  0.3962056821558659,
  1.5000000000000000,
  0.6109417595177361,
  0.6745091930459820,
  0.6128473721673339,
 -1.0168194384482883
]

Both representatives generalize less well to the held-out release set than to the training set. Reporting that gap prevents the favorable training score from being mistaken for general ventilation performance.

Straight-channel reference

The independent 48 × 20 full-height straight channel uses 1,200 allowed flow steps and converges after 980:

This verifies symmetry, low transverse flow, profile development, flux conservation, and residual convergence for a simple case. It is a numerical property test, not validation against measurements. Raw output is in verification/channel-reference.csv.

Three-grid sensitivity

The scalar horizon scales linearly with grid width and the flow limit scales approximately with width squared:

Reporting qualities:

The coarse-grid MODE representative has 5.326% interior flux mismatch against the 5% limit. Its scalar therefore contains a constraint penalty. The 40 × 24 and 60 × 36 evaluations are feasible, but their remaining quality differences show that the study establishes sensitivity rather than formal grid convergence. Raw data is in verification/resolution-study.csv.

Figures

All optimization figures use seed 42, while the summary table and right-hand verification panel retain all three seeds.

Artifact map

For every mode-seed-* directory:

• pareto.csv: feasible final non-dominated points and decisions;
• convergence.csv: search trace;
• validation.csv: selected design on held-out releases;
• selected-field.csv: final field for the worst-exposure training release.

For every qd-seed-* directory:

• archive.csv: occupied niches, descriptors, quality, visits, and decisions;
• convergence.csv: coverage, QD-score, and best-quality trace;
• validation.csv: best scalar elite on held-out releases;
• selected-field.csv: its worst-exposure training field.

verification/ contains the reference, resolution study, and consistently reproduced baseline/MODE/MAP-Elites fields used by the field figure.

Reproduction cautions

• Optimizer seeds control candidate generation; objective simulations are deterministic.
• A Rust objective evaluation includes three pollutant solves, so evaluation counts are directly comparable between MODE and MAP-Elites.
• Held-out and verification evaluations are additional and excluded from the optimizer budget.
• QD-score depends on archive capacity and quality scaling and should only be compared under the recorded configuration.
• Three seeds reveal variability but are too few for strong statistical claims.