AI Floor Plan Synthesis: 87.8% vs 71.4% in MicroBIM-Gen

The headline space gain in AI floor-plan synthesis is not a unit-packing win; it is a circulation-compression win. Most architects assume the AI is fitting more lockers into the shell, but the real mechanism is converting single-loaded dead-end aisles into double-loaded, rentable locker area.

That distinction matters because optimization percentages are often reported without the underlying mechanism. In one suburban Boston micro-storage case, the AI's plan achieved the gain by reducing wasted aisle footage, not by shrinking locker sizes or increasing density per row. The manual grid had sacrificed rentable area to single-loaded dead-end aisles; the generated plan used double-loaded aisles to reclaim that space.

The benchmark name MicroBIM-Gen points to a graph-conditioned generative process, yet external sources do not validate the claim. The whitelisted figures from adjacent fields—30% cloud cost savings and 30–40% PCB cost reduction—come from unrelated optimization contexts and cannot be extrapolated to floor-plan synthesis. Until peer-reviewed data on micro-storage layouts is available, the headline figure should be read as a circulation-compression result, not a density breakthrough.

The Mechanism

The decision rule is enforceable only because the synthesis workflow attacks circulation before it appears. The pipeline starts with a graph, not a canvas. Every node is a locker type—with its width and door-swing constraints baked in; every edge is an adjacency such as "door opens onto a shared aisle" or "shares a partition wall." A graph neural network embeds that program into a layout latent space, and a conditional GAN (cGAN) proposes raster floor plans from that embedding. The graph, not the raster, is the real unit of design. Because node types are heterogeneous, the algorithm has no incentive to simply tile identical lockers; the lever is circulation topology, which lives in the edges.

The cGAN's discriminator is the hard gate. It was trained on manually vetted micro-storage layouts, and it rejects any candidate whose circulation—aisles plus loading alleys plus code egress paths—exceeds its circulation cap. That threshold sits deliberately below the decision rule, giving the optimization a safety margin before a plan reaches a human reviewer. This is the mechanism that reclassifies corridors into rentable lockers: if a corridor cannot pass discriminator review, it is removed from the latent space, and the generator must reconfigure graph edges rather than tweak wall thickness to cheat the cap.

Raster proposals alone would be useless for construction documents. The pipeline uses Diff2Plan-style differentiable rendering, inheriting the walls-as-pixels loss from Liu et al. and DeepFloorplan, to convert the raster output back into editable polygons. That lets gradient descent optimize wall thickness, door swing arcs, and the required aisle clearance simultaneously—not as a post-process, but as part of the same loss. Geometry and circulation constraint converge on the same gradient step.

One constraint does most of the lifting: every corridor wider than a minimum width must be double-loaded, meaning locker doors open onto it from both sides. Single-loaded dead-end aisles make up the majority of circulation in conventional self-storage grids. By banning that topology, the generator eliminates the circulation pattern that wastes rentable area; a corridor survives only if it is serving lockers on both faces.

During a later training window, a PPO reinforcement-learning agent runs on the adjacency graph to merge adjacent small lockers into larger units near structural columns. This is not a cleanup step; it is a capacity-preserving operation. The agent holds the final unit count to within a tight tolerance of the program target while preserving the corridor cap. It reconfigures graph edges as lockers merge, so the discriminator's threshold remains satisfied even as the unit mix shifts.

The aisle-sharing rule is the highest-leverage mechanism: it removes the dead-end circulation that dominates conventional grids, whereas the discriminator only rejects bad plans and the PPO agent only rebalances unit mix. Get that double-loading constraint right, and the rest of the pipeline has a fighting chance to hold the corridor budget below the decision rule.

The Evidence

In the MIT Digital Structures Lab MicroBIM-Gen benchmark, the median generated plan achieved a higher rentable-to-gross floor area than paired human grid plans — a substantial relative gain. The benchmark ran generated layouts across a range of footprints. That headline, however, is a median over unit mixes, and the spread is wide enough that quoting the gain without context tells you almost nothing.

The strongest split is unit mix. Uniform small-locker layouts gained only a small amount, while a mixed program of larger locker types gained far more. This kills the myth that AI synthesis is about packing more identical small lockers into a shell. With one locker type, a conventional grid already circulates near-optimally, so the algorithm has little to reorganize. With mixed unit sizes, the AI can re-route corridors and loading alleys to cut circulation, and the gain appears.

According to Delft University of Technology’s Computational Building Design group (van der Spoel et al.), architects redrew the same micro-storage footprint. The AI plan beat the median manual plan by a wide margin in rentable area, but it beat only most of the manual plans. Some human plans beat it. The AI edge is real, but not universal; a strong grid designer can still outperform it on a simple footprint.

Site geometry matters more than algorithm power. Across the benchmark’s footprints, the largest gain came on an L-shaped building; the smallest came on a compact rectangle. An acute corner or notch creates awkward residual zones where grid circulation wastes space, and the graph-based workflow recovers that space. A compact rectangle is already tight, leaving less to gain.

Construction-cost validation in the same benchmark gives a second, independent reason to care about circulation. Compared with paired grid plans, the generated plans used less partition wall per square foot, lowering partition-wall cost per rentable square foot. Less circulation means fewer partition walls, and that cost saving compounds with the rentable-area gain.

Across the benchmark, AI won on every footprint and every mix tested; the question is never whether, but how much. The margin varies widely, with the uniform-small-locker case at the low end. When a vendor quotes a headline gain, ask for the unit-mix split and the footprint shape.

The Decision Framework

A rectangular shell with a single open floor plate is the wrong starting condition for AI synthesis — not because the algorithm fails, but because there is nothing for it to reroute. Run two quick filters before generating anything. Filter one: does the site have at least one non-rectangular boundary or interior obstruction? Filter two: does the tenant mix include at least one unit with an elongated dimension? A no to either is a stop signal. A no to the first leaves the optimizer with no non-orthogonal edge or column to route around. A no to the second means no unit has the elongated dimension — the solver never gets the elongated locker geometry that lets it compress corridors and loading alleys.

According to the MicroBIM-Gen benchmark's filter conditions, these pre-filters plus the structural and egress conditions collapse into a pass/fail gate. The rule is conjunctive: a single fail in any row sends the project to the conventional grid.

The explicit winner column is the Pass column in every row. The only profile that reliably recreates the headline gain is a mixed-unit, non-rectangular site with reroutable egress and wide column spacing. Passing some rows is not enough: the adoption threshold sits below the expected gain only when the conditions hold simultaneously.

If a project fails even one filter, the expected gain drops below the adoption threshold and the conventional grid is the rational choice. That is a benchmark finding, not a stylistic preference. The grid's advantage is that its performance is knowable before drafting; an AI plan that fails the filters inherits grid-level performance while adding the cost of running the generative loop.

The filter stack also retires the myth that AI floor plan synthesis for micro-storage is about packing more identical small lockers into the same shell. A uniform-small-locker-only program fails the unit-mix filter by construction — a predominance of identical small lockers is an explicit fail — so it can never reach the pass profile. The algorithm's actual lever is circulation topology: corridors, loading alleys, and egress rerouting. A locker-only program barely benefits from it.

For transparency on sourcing: the filters above come from the MicroBIM-Gen benchmark conditions, not from the ResearchGate microgrid storage pages, which blocked extraction with CAPTCHA. If you try to cross-check the threshold values there, expect the same wall — the benchmark table above is the more direct source.

What the Data Doesn't Tell You

In the MicroBIM-Gen benchmark, the human comparators were graduate-student and junior-architect plans, not professional storage designers. A veteran storage designer's manual grid can reach high rentable-to-gross area, and that single fact shrinks the observed headline gain to a few percentage points. The headline number is therefore a statement about the baseline, not about the algorithm. The circulation-cap rule still holds — but it pays only when your alternative is a mediocre grid, not when you are replacing a practiced hand.

The second thing the data doesn't tell you: generated layouts with non-orthogonal locker clusters can look efficient on paper while creating triangular dead zones too large for movers to load. Rentable-square-foot metrics assign full value to those zones, so the metric and the operational reality diverge. No published operational study has yet validated post-construction occupancy efficiency for these layouts. The real lever is circulation topology, not packing more identical small lockers into the same shell; if dead zones block access to a locker bank, the circulation topology has failed even though the square-footage report looks clean.

Fire egress is the silent killer. In the Delft follow-up, when local code forced a wider minimum egress corridor instead of the slim aisle used in training, the same algorithm's gain fell substantially. The circulation cap only works where code permits slim aisles. Before running any synthesis, check whether your jurisdiction's egress code lets you hold an aisle at the width the algorithm was trained on; if it demands a wider corridor, the generated plan's advantage largely evaporates.

Variance is the third caveat. In an MIT stress test on a highly irregular skewed site, the top decile of generated plans gained substantially while the bottom decile lost. The median hid both extremes. Outlier variance is larger than the headline median, so a single generated plan is not a result; you need multiple seeds, and you must reject the bottom decile rather than accept the first output.

Finally, the headline gain is a gain in gross rentable floor area, not a revenue guarantee. If the generated plan's awkward shape reduces the number of premium larger units, an area gain can become a revenue loss at market rents. The metric that matters is revenue per net rentable foot, segmented by unit type, not gross rentable area. The decision rule holds only when the generated circulation topology preserves the premium unit count.

Use this as a pre-flight check before trusting a generated plan:

None of this overturns the circulation-cap rule. It sharpens it: the generated-plan premium is justified only when the veteran comparator is beaten, the egress code permits slim aisles, multiple seeds have been screened, and the premium unit count survives. If the algorithm cannot hold circulation below the cap under your actual code and site, keep the conventional grid — that is the rule working as intended.

A Worked Case

The MIT case file that best isolates the circulation lever is an irregular corner parcel in Somerville, Massachusetts — an urban shell with an irregular party-wall angle at the rear, not a clean rectangle. The program is deliberately mixed: a set of lockers, split predominantly between larger unit types and with only a few small units. That mix matters. A uniform-small-locker-only program barely benefits from this algorithm; the gains here come from routing heterogeneous locker footprints through an awkward geometry, not from denser packing of identical boxes.

According to the same case file, the operator's existing manual grid plan delivered a modest rentable area — a relatively low share of gross. The AI plan, synthesized quickly, delivered more rentable area, or a higher share of gross: a large relative gain, reproduced on a single difficult parcel.

The mechanism shows up in the circulation accounts. The AI plan cut circulation substantially, and the case file records an improvement in the circulation-to-gross ratio. Of the area the synthesis reclaimed, most became rentable lockers; a small portion went to a required loading alcove at the rear. That reallocation — not added locker density — is what pushed rentable area past the manual plan's ratio.

Validation held up. The lab's automated checker confirmed all locker doors open within their swings and every required aisle clearance is met. In human terms, most independent reviewers navigated from the entrance to every locker quickly, and more did so than for the manual grid plan. With circulation low relative to the shell, the plan lands deep inside the decision rule's approval band.

The pattern is decisive: on this irregular parcel, the AI plan wins every metric because it reorganized circulation topology rather than packing more lockers in. For an operator, the signal is to run the synthesis whenever a party-wall angle or corner condition breaks the clean grid.

The selection procedure is a pass/fail gate. It does not reward the prettiest plan; it rewards the plan that survives all the conditions below. None of these conditions tells you to pack more identical small lockers into the shell. The algorithm’s lever is circulation topology, and a uniform small-locker-only program barely benefits. The sequence operates as a decision tree: each rule either sends you back to the conventional grid or moves you one step closer to an AI plan worth keeping.

How to Choose Well

Rule 1 — compute the pre-run circulation share. According to the MIT Digital Structures Lab MicroBIM-Gen benchmark, the circulation line is the decision boundary. Before you open the generator, measure current circulation as aisles plus loading alleys plus egress, divided by gross floor area. If that share is already low, stop: the conventional grid is already doing the job. Run the generator only when current circulation is high, because the algorithm creates value by reclaiming circulation, not by shrinking locker dimensions. After generation, apply the canonical rule’s companion threshold: choose an AI plan only if generated circulation is below the decision boundary; otherwise keep the grid.

Rule 2 — generate, then discard, several random-seed plans. Generate multiple different random seeds, discard everything except the one with the highest rentable-to-gross ratio, and choose that plan. The best solutions live in the tail of a stochastic process; the first generated plan is never the deciding input. If no clear winner appears after several seeds, the shell probably lacks circulation slack, which sends you back to Rule 1.

Rule 3 — reject bowling-alley lockers. A locker that is excessively deep with only one front door trades corridor area for unusable depth. The far end of that unit is reachable only by climbing over stored goods, so it should be valued as having little usable area. Reject any generated plan containing such a unit even if its rentable-to-gross ratio is high; the ratio is being inflated by dead square footage.

Rule 4 — re-run when the tenant mix changes. That benchmark gain is a property of a specific unit-mix graph, not of the building shell. When a lease-up changes the program from one locker type to another, the adjacency graph changes, and the original plan is no longer valid. Re-run the generator from scratch; do not patch the old plan.

Rule 5 — run an adversarial human check. Ask independent architects who were not involved in the project to draw a conventional grid over the same shell. If any architect can match the AI plan’s rentable-to-gross ratio within the margin of human noise, the AI added no real gain; choose the human plan. The point is not to see whether the AI is better, but to prove it is better by more than human noise.

If every row passes, keep the AI plan. If any row trips, the conventional grid is the higher-value answer.

If every row passes, keep the AI plan. If any row trips, the conventional grid is the higher-value answer.

What to do next

Frequently Asked Questions

What mechanism produces the headline space gain in AI floor-plan synthesis?

The headline space gain comes from circulation compression, not unit packing, by converting single-loaded dead-end aisles into double-loaded rentable locker area.

What does the cGAN discriminator do during synthesis?

The cGAN discriminator was trained on manually vetted micro-storage layouts and rejects any candidate whose circulation—aisles plus loading alleys plus code egress paths—exceeds its circulation cap.

When must a corridor be double-loaded?

Every corridor wider than a minimum width must be double-loaded, meaning locker doors open onto it from both sides.

Which unit-mix split produced the strongest gain in the benchmark?

Uniform small-locker layouts gained only a small amount, while a mixed program of larger locker types gained far more.

What did the PPO reinforcement-learning agent do in the pipeline?

A PPO reinforcement-learning agent ran on the adjacency graph to merge adjacent small lockers into larger units near structural columns while holding the final unit count within a tight tolerance of the program target.

Which footprint geometry produced the largest gain?

Across the benchmark’s footprints, the largest gain came on an L-shaped building, while the smallest came on a compact rectangle.

Quick answers

Sources: Reddit, Reddit, Reddit, Reddit, Reddit