Doklad

1.

A Method of Crystal Lattice Recognition using
AI/ML. A case study of Uranium Carbides and
Nitrides.
Sergey S. Pitskhelaury1*, Ivan D. Burdukov1, Artem A. Artiushin1,
Alyona D. Markova1, Matvey S. Chernyakov1, and Kirill A. Nekrasov1
1) Ural Federal University named after the first
President of Russia B.N. Yeltsin, Yekaterinburg, Russia
* s.s.pitskhelauri@urfu.ru
INTRODUCTION
ON THE APPLICABILITY OF THE POINTNET ARCHITECTURE
Uranium carbides and nitrides are promising fuels for Generation IV reactors, but their
industrial use, unlike MOX fuel, requires deeper insight into vacancies, cascades, and
mixed carbonitride behavior—issues still unresolved. Experiments show that at melting
temperatures, alternative-phase nuclei (U₂C₃, UC₂, U₂N₃, UN₂) emerge in monocrystals,
yet tracking this in real time is experimentally unfeasible, so molecular dynamics
simulations are used. However, standard algorithms fail to process the resulting data
deluge effectively, necessitating AI/ML integration. The goal is thus to build a data
processing framework that automatically detects and tracks nucleation events in
simulation outputs for this material.
b)
a)
Fig. 2 — Demonstration of the PointNet Architecture in Operation:
a) Schematic illustration of the data processing workflow;
b) QR code providing access to the conceptual video
a) UC
b) -UC2
c) -UC2
d) U2C3
Fig. 1 — Illustration of uranium carbide crystal geometry.
PROPOSED CONCEPT
Fig. 3 — Conceptual model of the proposed solution.
The proposed algorithm operates in two stages. First, molecular dynamics simulation is performed, and every 100 time steps, data on ion positions are saved for subsequent
processing by the PointNet architecture, which performs spatial segmentation of crystal structures and generates phase-affiliation hypotheses. These hypotheses are then passed to the
KDE module for statistical verification (using Random Forest and CatBoost methods for recognition against reference plots stored in the database), after which a final report is
issued. The two-stage design is driven by the mutual limitations of the methods: PointNet effectively localizes regions but is sensitive to reference shapes and produces errors at phase
boundaries; KDE provides a reliable probabilistic assessment but fails to determine the spatial location of nuclei and yields false positives under high defect densities. Their
combination compensates for each other's weaknesses, and the final mechanical verification additionally filters out false results.
MATHEMATICAL MODEL OF KDE-BASED ANALYTICS FOR PHASE VERIFICATION
Step 1. For each ion, a sample of distances to the N nearest neighbors is constructed: