Features

STL File I/O

• Load ASCII or binary STL files from disk via surface%load_from_file
• Automatic format detection (guess_format=.true.) — no need to know the file format upfront
• Load with on-the-fly clipping: clip_min / clip_max bounding box arguments discard facets outside the box during load
• NaN/Inf coordinates are rejected at load time via the optional status argument (STATUS_INVALID_INPUT); garbage input never enters the geometry pipeline
• Save surfaces to ASCII or binary STL via surface%save_into_file

Surface Analysis

After calling surface%analyze (or surface%sanitize, which calls it internally), the following data are computed and stored:

• Axis-aligned bounding box (get_bmin, get_bmax)
• Enclosed signed volume (get_volume) — positive for outward-oriented closed bodies (standard divergence-theorem convention) and centroid (get_centroid)
• Facet connectivity — symmetric edge-adjacency via the sort-and-pair algorithm (matches trimesh / Open3D / libigl). f1.fcon_edge(e) = f2 always implies a back-link from f2 to f1
• Non-manifold edge detection (3+ incident facets are explicitly counted, never silently collapsed to a single neighbour)
• Disconnected-edge detection: facets with 1, 2, or 3 disconnected edges are catalogued separately
• AABB acceleration structure (SAH BVH by default; octree available)
• Vertex / edge angle-weighted pseudo-normals (Bærentzen–Aanæs) used by signed-distance queries

Validity predicates

Predicate	Returns .true. iff
is_watertight()	Every edge has exactly two incident facets (no boundary, no non-manifold)
is_manifold()	No non-manifold edges (boundary edges allowed — open shells can be manifold)
is_volume()	is_watertight() AND volume > 0 AND finite centroid

Surface Manipulation

All manipulation methods operate directly on surface_stl_object:

• Translate — by a 3D vector (delta=) or by scalar components (x=, y=, z=)
• Rotate — around an arbitrary axis by angle in radians, or by a given rotation matrix
• Mirror — with respect to a plane defined by its normal, or by a given mirror matrix
• Resize (scale) — by a 3D vector factor or by scalar per-axis factors; optionally scale about the surface centroid
• Clip — discard facets outside an axis-aligned bounding box; the cut-off remainder is returned as a separate surface
• Merge — combine two STL surfaces into one (merge_solids)

Mesh repair pipeline

surface%sanitize runs the full repair sequence in the right order, with one stderr warning per defect class detected:

1. remove_degenerate_facets — drop zero-area / sliver triangles before their NaN normals contaminate downstream queries
2. connect_nearby_vertices — union-find vertex deduplication on a tolerance
3. analyze — rebuild metrix, connectivity, vertex occurrences
4. remove_duplicate_facets — drop literal duplicates (orientation-agnostic; matches the typical CAD-export defect)
5. analyze — rebuild after duplicate removal
6. sanitize_normals — BFS-propagate winding consistency across the connectivity graph, then flip globally so volume > 0 (outward orientation)

Individual passes (remove_degenerate_facets, remove_duplicate_facets, connect_nearby_vertices, sanitize_normals, reverse_normals) remain callable for fine-grained control.

Distance and point-in-polyhedron queries

• surface%distance(point, is_signed, is_square_root, sign_algorithm) returns the minimum distance from a point to the surface
  • Default returns squared distance; pass is_square_root=.true. for Euclidean
  • Pass is_signed=.true. for the signed distance (negative inside, positive outside)
• Sign-determination algorithms (sign_algorithm argument):
  • SIGN_PSEUDO_NORMAL (default) — Bærentzen–Aanæs pseudo-normal test. Fused with the closest-facet traversal so signed distance costs the same as unsigned. Requires consistently oriented outward normals (which sanitize produces).
  • SIGN_RAY_INTERSECTIONS — odd/even axis-aligned ray-cast count. Robust on meshes with mixed orientation but ~3× slower than pseudo-normal.
  • SIGN_SOLID_ANGLE — sum of projected solid angles ≈ ±4π. Robust but unaccelerated; O(N) per query.
• surface%is_point_inside(point, sign_algorithm) is the predicate form of the same test.

All tree-accelerated queries are bit-exact vs. brute force on the regression suite — best-first traversal with d² pruning never prunes the true closest facet.

AABB acceleration

Two tree structures are available; both share the same traversal code and produce identical query results:

• AABB_TREE_SAH_BVH (default) — binary BVH built top-down by partitioning triangles along the longest centroid-bbox axis, choosing splits via the bucketed surface-area heuristic (16 buckets, 3 axes, evaluate at each bucket boundary). Adapts to local triangle density.
• AABB_TREE_OCTREE — 8-way space-partitioning octree built to a uniform refinement depth (auto-tuned from facet count via AABB_AUTO_REFINEMENT).

Selection is per-surface via the optional aabb_tree_kind argument on load_from_file / adopt_facets / analyze. Benchmark numbers (gfortran -O2, dragon-fine, 24318 facets, 32³ query points):

Tree	Build	Query
Octree	0.53 s	27.32 s
SAH BVH	0.32 s	0.25 s

fossilizer CLI

A command-line utility (src/app/fossilizer.f90) wraps the library for interactive STL processing:

• Load one or more STL files
• Apply clip, merge, mirror, rotate, translate, resize, sanitize, and connectivity repair operations
• Print per-file statistics to stdout
• Save the result to a named output file

Compiler Support

Compiler	Status
GNU gfortran ≥ 5.3	Supported
Intel Fortran ≥ 16.x	Supported
NVIDIA HPC Fortran (nvfortran)	Supported
IBM XL Fortran	Not tested
g95	Not tested
NAG Fortran	Not tested

Design Principles

• Pure Fortran — no C extensions, no system calls beyond standard I/O
• OOP — all functionality exposed as type-bound procedures; the public API is just surface_stl_object and facet_object
• TDD — every public operation is exercised by automated tests in src/tests/. The signed-distance regression test asserts bit-exact agreement between tree-accelerated and brute-force queries
• KISS — simple, focused API without unnecessary abstractions
• Free & Open Source — multi-licensed for both FOSS and commercial use