∴Materials Growth LabLOCAL OBSERVATION → STRUCTURAL CONTINUATION
GCTS method ↗Research console ↗
THE ATOMIC PATCH IS THE STARTING POINT

Learn the pieces.
Explore what can grow.

Species and coordinates become recurring clusters, oriented connections, and an expanding search tree.

ACTIVE SPECIMEN

NaCl

KNOWN ATOMIC POSITIONS

The observation

element colors retainednew atoms softly illuminateddrag to orbit · scroll to zoom
EXPLICIT ATOMS216
CLUSTER TYPES—
OVERLAP PORTS—
BACKTRACKS—

Choose a specimen or proceed with NaCl.

THE SEARCH, MADE VISIBLE

A decision is a branch

  1. Observed atoms remain fixed throughout search.
  2. Learned ports place whole clusters by rotation and translation.
  3. A failed continuation rolls back to a retained alternative.
HOW TO READ THE EXPERIMENT

Geometry is a hypothesis

Overlap and collision checks certify compatibility with the learned geometric rules. They do not establish phase stability, an energy minimum, or physical growth rates.

Compare saved markings using explicit growth, checks and backtracks. A low training loss alone is not evidence of better continuation.

Open the physics and benchmark research console ↗
AUDIT · SEPTEMBER 2026

What is implemented?

The workspace follows the reversible search structure of GCTS-I: enumerate, check, order, place, and undo. Search begins with the entire observed patch.

Continuous geometry
Colored supports are compared under proper SE(3) at the declared tolerance. Candidate poses come from witnessed overlaps and cluster symmetries; no spatial lattice is imposed.
Coverage
Repeated supports and residual atoms cover the finite observation. This covers atoms, not empty volume. The learner does not assume periodic images.
Molecular supports
When repeated isolated components dominate the patch, supports preserve whole components. Unrecognized crop fragments remain seed residuals. Molecular closure is inferred from geometry, not a chemical bond model.
Markings
The channel model learns observed connection frequencies and ranks legal choices. Its loss and halos are training diagnostics.
Online agent
The experimental bandit learns port values from novel-site rewards and rollback penalties. It is not a pretrained RL policy or physical-quality estimator.
Hierarchy
Optional repeated unions of two overlapping clusters require atom-disjoint witnesses. One promotion level is supported; stationary recursion is not established.
Computation
A worker runs search with atom, candidate and branch memory limits. The best structure survives rollback. Exhausting a finite catalogue does not prove that no larger material exists.

Read the full implementation audit ↗