Can Polat pronounced “John”
2026 · Preprint

Physical survival and provenance limits on watermarking AI-generated crystals

Can Polat , Mustafa Kurban , Erchin Serpedin , Hasan Kurban

Research highlights

BasinMark: auditing AI-generated crystals after physical relaxation

Generated crystals undergo relaxation and standardization before researchers use them. Those steps can erase the structural signatures intended to identify their origin. This makes provenance a practical question for anyone building or reusing datasets of AI-generated materials: what evidence survives the actual scientific workflow?

Why we investigated it

A watermark detected immediately after generation may disappear during physical processing. Repeated or equivalent structures can also make an audit appear more convincing than its independent evidence supports. We study these effects together so researchers can judge both the survival of a watermark and the meaning of a positive detection.

What the work contributes

BasinMark selects generated candidates using keyed structural fingerprints and audits the structures that remain after processing. The paper identifies which fingerprint features are sensitive to relaxation and derives false-positive control by counting each distinct fingerprint class once, under the stated random-function and key-independence assumptions.

What the evidence shows

Across the evaluated generators and interatomic potentials, cell-shape bins account for much of the observed fingerprint instability. Removing them improves label survival in two generator populations while retaining nearly all observed fingerprint classes. The reuse experiments also establish an important boundary: a positive audit identifies a marking signal at the fingerprint-class level; it cannot uniquely authenticate the origin of an individual crystal.

Build on this work

Use the repository to compare fingerprint designs, evaluate sensitivity to relaxation, and examine how structural duplicates affect audit confidence. It includes analysis code, reported numerical summaries and tables, and Lean formalizations of mathematical components with an explicit coverage map. Structure-based experiments require users to supply the input structures and manifests.

  • Crystal watermarking
  • AI-generated materials
  • Materials provenance
  • Structural fingerprints
  • Crystal relaxation
  • Statistical auditing
  • Generative materials design
Abstract
When crystals are generated, they are relaxed and standardised before scientific use. However, this can erase or alter structural evidence of their production. We ask what a structure-only watermark can preserve through this processing, how an auditor should calibrate the surviving evidence, and which provenance claims the resulting decision supports. BasinMark connects the deposited observation to a collection audit by selecting generated candidates according to keyed structural fingerprints. Exact relaxation to the same basin minimum erases coordinate marks, so the protocol instead selects among alternatives whose fingerprint distinctions may survive deposition. Across the evaluated generators and interatomic potentials, cell-shape bins account for much of the observed fingerprint instability; removing them in two generator populations substantially improves fingerprint-label survival while retaining nearly all observed fingerprint classes. Under an ideal random-function model and a key-independent null, counting each unique fingerprint class once provides exact false-positive control. However, because detection operates at the structural class level, collections with identical fingerprints remain indistinguishable—allowing watermark evidence to be copied or reused across datasets. Ultimately, physical processing, fingerprint choices, and structural duplicate handling determine how much watermark evidence actually survives.
Cite
@misc{polat2026watermarking,
  title = {Physical survival and provenance limits on watermarking AI-generated crystals},
  author = {Can Polat and Mustafa Kurban and Erchin Serpedin and Hasan Kurban},
  year = 2026,
  url = {https://johnpolat.com/publications/polat-2026-crystal-watermarking/paper.pdf},
}