DEV Community

GeoGenius
GeoGenius

Posted on

Landform as Data: GenAnywhere's Entity Model for Surfaces, Benches, and Cut-Fill Balance

Terrain data in most toolchains is a surface and nothing else: a triangulated mesh that knows elevations but not intentions. Ask it why a bench exists at a given height, or where the water on a given face is supposed to go, and it returns silence, because those facts were never modeled. GenAnywhere takes the position that a landform is not a byproduct of a surface but a first-class entity with a type, a purpose, and obligations — and that the surface is merely the rendered consequence of the entities placed upon it.

The landform entity vocabulary

The model begins with a Surface entity holding the existing ground as an immutable baseline, and a set of Landform entities that describe intended changes. A Bench is a flat cut or fill step carrying an elevation, a width, and a face slope. A Berm is a shaped rise carrying a crest elevation and a screening or holding purpose. A Swale is a shallow conveyance carrying a grade and an outlet reference. A Terrace is a composite: an ordered collection of Benches sharing a toe strategy. A Pad is a working surface carrying a finish elevation and a tolerance. A Channel is a defined conveyance with a section and a capacity.

Typing is what makes these entities enforceable. A Swale cannot exist without an outlet reference, because a conveyance to nowhere is a defect by definition. A Bench on a Terrace inherits the terrace's toe strategy, so changing the strategy updates every bench at once. A Pad declares the landforms that shed to it, which is how a working surface accumulates its drainage obligations automatically rather than by memo.

Drainage invariants as the core intelligence

The intelligence of the model lives in invariants that bind landforms together. The continuity invariant requires that every Swale and Channel terminate in a named outlet or a Pad that accepts flow, and it re-evaluates on every edit; sever a swale from its outlet and the edit is rejected with the invariant named. The capture invariant requires that each graded face shed to exactly one declared conveyance, preventing the ambiguous shedding that produces mystery ponding. The capacity invariant compares conveyed volume against channel section on storm events, flagging any reach that overtops.

A second family governs earth. The balance invariant tracks cut and fill across the project as landforms are added, so a team sees the net balance move in real time instead of discovering it in a quantity takeoff weeks later. The benching invariant requires that any fill placed over ground steeper than a threshold declare benching steps, encoding a geotechnical rule directly into the model. The settlement invariant attaches tolerance constraints to landforms placed over capped or compressible ground, which is how a reclaimed landfill or a backfilled pit carries its restrictions forward automatically.

Surfaces as rendered consequences

Given the entity set, GenAnywhere derives the finish surface deterministically. Benches and terraces cut or fill the baseline; swales and channels incise their conveyances; berms add mass. Because derivation is deterministic, two states of the entity set can be differenced into a surface change map: here the ground rose by a meter, here it fell, here a conveyance moved. Review becomes a conversation about deltas rather than about two opaque meshes, and a design amendment produces an exact before-and-after pair at any station.

The same derivation feeds the asset pipeline. A terrain state renders at three scales: the site aerial for orientation, the slope-section view for engineering review, and the ground-level vignette for public engagement. All three are projections of one entity set, so the council packet, the geotechnical review, and the community meeting are looking at geometrically identical ground. Nothing drifts out of sync because there is only one source of truth.

Landform purpose and the memory problem

Every Landform entity carries a purpose field, and purpose is load-bearing. A Terrace whose purpose is inundation sequencing behaves differently under review than a Terrace whose purpose is seating: the former is checked against flood stages, the latter against accessibility slopes. Encoding purpose means the correct checks attach themselves, and it means the reason a bench sits at a given elevation survives staff turnover. The model treats forgotten intent as a data-integrity failure, not as an inevitability.

Agricultural terraces illustrate the point economically. A citrus terrace exists because cold air drains downhill and the riser holds a warmer band; that purpose implies frost-drainage swales at each toe and a minimum face slope for air movement. Encode the purpose once and the implied geometry checks itself on every edit. Leave purpose implicit and each engineer re-derives the cold-air story from scratch, occasionally getting it wrong in a way that only surfaces on a freezing night years later.

Why the model changes behavior

Teams report that the invariant-first model changes when mistakes surface: at the keyboard, seconds after introduction, rather than at construction or in the first storm. It also changes what teams argue about. With continuity, capture, capacity, balance, and benching enforced automatically, the remaining disagreements are genuinely design questions — where the bench should be, how steep the face, which hollow becomes wetland — instead of bookkeeping disputes about whether the water has somewhere to go. That is the practical payoff of treating landform as data: the machine absorbs the bookkeeping, and the humans get back the design.

Top comments (0)