The mesh beneath the site

Before a single pile is driven, before a geotechnical report lands on the project server, there is a system already running beneath every building site on earth — and until very recently, no one had a map of it. As Quanta Magazine reported in April 2026, the nonprofit Society for the Protection of Underground Networks (SPUN) published a machine-learning model in Nature in 2025 that processed 25,000 geolocated soil samples and generated more than 2.8 billion fungal DNA sequences to locate global hot spots of mycorrhizal fungi. The result: the first quantitative biodiversity atlas of the networks that quietly govern soil stability, carbon storage, and plant-ecosystem health worldwide.

For architects and civil engineers, the instinct is to file this under “interesting biology, not my problem.” That instinct is wrong.

←TODAY: SPUN’s 2025 Nature model identifies Arctic Alaska as a mycorrhizal hot spot using ML inference from 25,000 soil samples — the first planetary-scale map of this underground infrastructure layer.
→3012: Every site masterplan carries a subsurface biodiversity model alongside the geotechnical log; foundation strategy and carbon accounting are co-optimised against living soil systems.
Fulcrum: The moment a biological network becomes a mapped, queryable dataset, it enters the design system — and the liability register.

Why the system is newly legible

Mycorrhizal fungi — an evolutionarily dispersed group of soil-dwelling microbes, likely comprising 20,000–50,000 species — were classified as plant parasites in the late 19th century, then reframed as passive infrastructure, and are now understood, per SPUN chief scientist Toby Kiers of Vrije Universiteit Amsterdam, as “active merchants” that direct nutrient flow and restructure soil chemistry across entire ecosystems. That reframing matters to anyone who compresses or chemically treats soil for a foundation. Hyphae networks extend horizontally dozens of feet from a root node; they are not a local effect.

What made SPUN’s Nature model possible in 2025 and not 2015 is a convergence of three inputs: cheap environmental DNA sequencing, ML inference capable of extrapolating species distributions from sparse samples, and a global soil-sample collection effort that has been quietly running for a decade. The same logic — instrument sparsely, infer densely — is how structural health monitoring systems work on bridges, how ground-penetrating radar is used in urban archaeology, and how BIM federated models are increasingly populated from point clouds. The data architecture is familiar. Only the domain is new.

What this means on a working desk this week

Three concrete intersections with AEC practice:

• EIA scope creep: Environmental impact assessments in Switzerland (under the Umweltverträglichkeitsprüfung framework) and Germany’s UVPG are increasingly expected to address soil biodiversity, not just surface vegetation. A quantitative mycorrhizal hot-spot layer — now technically producible using SPUN’s methodology — is a foreseeable addition to site assessments within the next planning cycle revision.
• Carbon accounting in ground works: Mycorrhizal networks are significant carbon sinks. Soil disturbance during excavation releases sequestered carbon. As Scope 3 emissions accounting tightens under the EU’s Corporate Sustainability Reporting Directive (CSRD, mandatory for large firms from fiscal year 2024), ground-disturbing works will face pressure to quantify biological carbon loss alongside embodied carbon in concrete and steel.
• Foundation typology decisions: The geotechnical instinct is to treat soil as a substrate to be stabilised or displaced. But if a mycorrhizal network is an active regulator of soil structure — as the latest robotics-and-imaging studies cited in the Quanta piece confirm — then disrupting it may degrade long-term soil bearing performance in ways that are not captured by standard CBR or SPT test results.

None of this is speculative in the sense of being far off. The data layer exists. The regulatory pressure is building. The missing link is that most AEC teams have no workflow — no IFC class, no BEP chapter, no Grasshopper component — for subsurface biological information.

Atelier: In PAZ’s →3012 framing, the site model is a four-layer stack: atmosphere, built fabric, geology, and living soil systems. SPUN’s ML methodology is the first tool that makes the fourth layer queryable at project scale — treat it as a new data source category alongside LiDAR and BIM, not as a footnote in an ecological appendix.

The failure mode worth naming

The risk is not that AEC professionals ignore this. The risk is that it gets absorbed into compliance theatre: a biodiversity checkbox ticked by a junior consultant using a coarse national-scale dataset, with no feedback into foundation design, no carbon delta calculated, and no site-specific sampling. SPUN’s model is a global inference tool; it identifies where to look, not what is actually there. Van Nuland’s team drove to northern Alaska with steel core tubes precisely because the model cannot replace ground-truth sampling. The same discipline applies to a development site in Basel or outside Graz.

Pull the SPUN dataset via their open data portal, overlay it on your next large-footprint site, and bring the output to the geotechnical engineer at the first ground investigation meeting — not the planning submission.

Source: Quanta Magazine