mathematics//computational geometry//isosurface
An isosurface is the surface made of all the points of a three-dimensional field where the field takes one chosen value, and it is the standard way to turn volumetric data (a CT scan of a casting, an MRI, a simulated temperature or pressure field, a density predicted by a neural network) into a surface that can be seen, measured or printed. It is the 3D version of a contour line on a map: the contour joins points of equal altitude, the isosurface joins points of equal density or temperature.
An isosurface is the surface made of all the points of a three-dimensional field where the field takes one chosen value, and it is the standard way to turn volumetric data (a CT scan of a casting, an MRI, a simulated temperature or pressure field, a density predicted by a neural network) into a surface that can be seen, measured or printed. It is the 3D version of a contour line on a map: the contour joins points of equal altitude, the isosurface joins points of equal density or temperature.
Sc={(x,y,z): f(x,y,z)=c}S_c=\{(x,y,z):\ f(x,y,z)=c\}Sc={(x,y,z): f(x,y,z)=c}
The field fff gives a number at every point of space and ccc is the chosen level: the bone threshold in a CT scan, 600 °C in a simulated furnace, zero for a signed distance field whose sign says inside or outside. The classic algorithm, marching cubes (1987), walks the grid of samples one cube at a time, checks which corners are above and below ccc, and places triangles where the surface must cross, so the output is a polygon mesh.
An isosurface is only as good as the grid and the threshold behind it.
The detail it can show is bounded by the sample spacing, and the shape it shows depends on the chosen level: move the threshold of a CT scan of a cast part by a few percent and a porosity defect appears or vanishes.
The meshes it produces are dense and uniform, with many thin triangles and a face count set by the grid rather than by the shape, and they carry staircase artefacts where the grid is coarse. They are faithful to the data and awkward to edit, which is why generative methods that write meshes directly were developed (3D generation).
Control over detail comes from the grid: refining it everywhere multiplies memory by eight per halving of the spacing, so adaptive grids (finer where the surface curves) are used for large volumes.
In industry it is the inspection workflow for industrial computed tomography: a casting or an additively manufactured part is scanned, its surface extracted, and compared against the CAD model to map wall thickness and internal voids.
Many neural 3D methods predict an implicit field (occupancy or signed distance) and rely on marching cubes at the end, so their visual quality is partly the quality of this last step.