AI-Generated Buildings Reveal How Design Choices Skew Material Stock Estimates

Two researchers at the University of Hong Kong have built tens of thousands of synthetic apartment buildings with artificial intelligence and parametric modeling to answer a deceptively simple question: how wrong can the standard method for estimating the materials locked inside our buildings actually be? Their study, published in the Journal of Industrial Ecology, shows that even when two residential buildings have exactly the same floor area, the amount of material embedded in their walls, columns, beams, and slabs can differ by nearly a factor of four, depending entirely on architectural design decisions that conventional estimation tools simply ignore.

The method at the center of the debate is the material intensity approach, the workhorse of bottom-up building material stock analysis. It works by multiplying a building’s gross floor area by an aggregated coefficient, expressed in kilograms or cubic meters of material per square meter of floor space. That simplicity is precisely why it dominates the field: it requires no component-level data, no floor plans, and no knowledge of internal layouts, making it feasible to estimate material stocks for entire cities, nations, or the global building stock. Those estimates matter enormously for the circular economy, because knowing what materials sit where in the built environment is the prerequisite for urban mining, the strategic recovery of steel, concrete, and timber from buildings at the end of their lives.

The problem, as Yingqi Jia and Chen Feng point out, is that the material intensity method treats all buildings of a given type as essentially interchangeable. Yet real buildings of the same class, built in the same era, can differ dramatically in footprint shape, corridor arrangement, number of units, room layouts, number of floors, and structural system configuration. Quantifying how much these design choices matter has long been blocked by a practical obstacle: detailed internal layout data for real buildings is scarce, and no city keeps a comprehensive database of floor plans suitable for this kind of controlled comparison. The researchers’ solution was to sidestep real-world data entirely and generate their own.

Their pipeline is a hybrid of two computational approaches, each compensating for the other’s weaknesses. Parametric modeling offers precise, systematic control over design parameters, which is exactly what a controlled experiment requires, but achieving architectural realism with parameters alone demands an impractically complex rule set. Deep learning generative models, by contrast, can produce realistic floorplans with minimal manual effort, but they do not allow explicit, systematic control of the variables under study. The team’s workflow combines both: parametric routines generate footprints, insert corridors, and subdivide space into units using a corridor-aware binary space partitioning algorithm, while a neural network called Graph2Plan, pre-trained on real-world floorplans, generates plausible room layouts within each unit. The pipeline then assembles three-dimensional structural components, resolves intersecting volumes with Boolean operations to avoid double counting, and calculates the material volume of every wall, column, beam, and slab.

Running this pipeline across the full parameter space produced 48,600 synthetic residential buildings, all with a constant gross floor area of 2,400 square meters, while systematically varying footprint shape among square, rectangle, L-shape, T-shape, and H-shape geometries; corridor type among single-loaded, double-loaded, and point-access systems; unit counts; room layouts; floor counts from one to six; and structural grid spacing and member sizing. The headline result is striking: total material stock ranged from 758 to 2,888 cubic meters across this design space, with a mean of 1,548 cubic meters and a standard deviation of 320. Under the traditional material intensity method, every one of those buildings would have been assigned the identical stock value, because floor area alone determines the estimate.

Disentangling which design parameters drive this variability required statistical analysis rather than inspection of individual cases. A multiple regression model explained more than 95 percent of the variance in material stock and revealed a clear hierarchy of influence. The number of floors was the single most powerful factor, alone accounting for roughly one-third of the variation, followed by the sizing of the structural frame, then the number of units and the footprint shape. Grid spacing and floor height came next, while corridor type and room layout had much smaller, though still statistically significant, effects. Every parameter in the model was significant at the p < 0.001 level.

The physical mechanisms behind these rankings are intuitive once spelled out. Holding floor area constant, the number of floors controls a building’s slenderness, and every additional floor must be enclosed by walls. In the synthetic dataset, wall volume grew from 600 cubic meters in single-story buildings to 1,186 cubic meters in six-story buildings, driving a 55.2 percent increase in total stock, while the structural frame grew far more modestly. Structural sizing mattered even more sharply within the frame itself: moving from slim to bulky member dimensions raised total material volume by 35.1 percent, and widening the structural grid from 4.0 to 8.0 meters cut it by 10.5 percent by reducing the number of columns and beams needed. Footprint shape operated through the surface-to-volume ratio, with the sprawling H-shape demanding 12.1 percent more material than the compact square, almost entirely in wall components, since frame volumes stayed nearly constant across shapes. Adding units raised stock sublinearly, because each new partition needs proportionally less wall, and going from one to three bedrooms per unit added only 5.4 percent.

Because the dataset is synthetic, the authors took plausibility seriously. They reproduced wall and slab volumes from an open residential building information model within 4 percent, checked that the 2,400 square meter floor area falls within real-world ranges observed in OpenStreetMap records across 17 cities on six continents, and compared converted mass-based intensities against the RASMI global benchmark database, producing a median of 1,276.77 kilograms per square meter against a benchmark of 1,285.52. They also examined structural calculation records for two real Hong Kong apartment towers, Tin Sam Villa and Millennium Court, finding derived column-and-beam volumes of 225.1 and 226.2 cubic meters, both close to the synthetic mean of 222.2 and well inside the synthetic range. Supplementary tests at 1,200 and 4,800 square meters confirmed the findings are robust to the choice of fixed floor area.

The practical implications cut in two directions. For analysts, the results suggest that material intensity databases should expand their building typologies beyond use type and construction period to include footprint shape, number of floors, and structural configuration, the parameters that dominate design-induced variability. Where such classification is not feasible, the pipeline can generate exploratory adjustment factors, nudging baseline intensities upward for slender, complex buildings and downward for compact, low-rise ones. For designers, the study offers a menu of material-efficiency levers: reduce slenderness, widen structural grids, slim down frame members, minimize unit divisions and floor height, favor square or rectangular footprints, and use point-access cores instead of corridors. The authors caution that these levers must be weighed against structural, regulatory, economic, and functional requirements they did not model, and that their analysis was limited to reinforced concrete frames, excluded foundations, roofs, and mechanical systems, and explored a theoretical design space rather than the distribution of real buildings, so the reported range represents potential variability rather than expected real-world spread.

What may prove most durable about the work is the methodological framework itself. By fusing parametric modeling, generative AI, and component-based material accounting into a fully automated pipeline, the researchers have created an extensible tool for running what-if experiments on building design and resource demand at a scale impossible with real-world data. The complete workflow, the database of more than 48,000 building configurations, and tutorials for reproducing the results are openly available on GitHub. As cities race to turn their building stocks into material banks for a circular economy, the study delivers a clear warning and a clear opportunity: floor area alone cannot tell you what a building is made of, but the design decisions that shape it can now be measured, ranked, and, ultimately, engineered for less.

Subject of Research: Sensitivity of building material stock estimates to architectural design parameters

Article Title: Assessing the sensitivity of material-intensity-based building stock estimates to design parameters

Article References: Jia, Y., & Feng, C. (2026). Assessing the sensitivity of material-intensity-based building stock estimates to design parameters. Journal of Industrial Ecology, 30(4), 2035-2050. https://doi.org/10.1007/s44498-026-00138-5

Image Credits: AI Generated

DOI: 10.1007/s44498-026-00138-5

Keywords: building material stock, material intensity, circular economy, urban mining, parametric modeling, deep learning, floorplan generation, structural design, reinforced concrete, industrial ecology, synthetic data, material efficiency