Role in the Project and Objectives (Materials Informatics)

Work package WP4 serves as the analytical and computational hub of the “GaN-Unipress” project. Its overarching mission is to transform heterogeneous technological and laboratory knowledge into standardized, machine learning-ready datasets (AI-ready data), while building predictive modeling tools that support packages WP1–WP3.

Activities within WP4 are organized into 5 pillars:

1. Data Analytics and Selection: Curating datasets (including rich historical ammonothermal growth records) that address critical process optimization questions.
2. Data Standardization and Ingestion: Converting disparate run logs, measurements, and micrographs into machine-readable tabular and tensor formats.
3. Data Repository: Establishing a central data management system ensuring full traceability, searchability, and genealogical tracking of specimens.
4. Predictive Modeling (AI/ML): Mapping physical process parameters to crystal morphology, defect emergence, and electrical performance.
5. Digital Twin: Constructing virtual multiscale simulations of autoclave growth to dramatically reduce empirical trial-and-error iterations.

The “Digital Crystal” Concept

The cornerstone of the WP4 architecture is the creation of a Digital Crystal object linked to a persistent identifier (ID), unifying:

• Computer Vision Descriptors: Contours, surface area, estimated volume, seed boundaries, lateral growth zones, and macroscopic defects.
• Structural and Physical Metrics: High-resolution XRD/XRT maps, rocking curve widths, optical transmittance, and electrical characteristics.
• Process History: Autoclave run ID, spatial basket position, pressure-temperature time-series, and genealogical ancestry (parent–child lineage between seed generations).
• CFD Simulation Features: Spatially resolved temperature profiles, buoyancy-driven flow velocity fields, mass transport rates, and supersaturation levels in the supercritical fluid.

Fig. 1. Architecture of the ammonothermal digital twin: integrating experimental datasets, digital crystal objects, CFD multiscale modeling, and reduced-order AI/ML surrogates.

Computer Vision and AI Tools for the Pilot Line

To support the high-throughput requirements of the WP1 pilot crystallization line, automated inspection pipelines are being developed:

• Hybrid Vision Approach (OpenCV + ML): Fast, deterministic computer vision algorithms (Python / OpenCV) handle camera calibration, contour extraction, and area quantification ($px \to mm^2$), while deep neural networks classify subtle surface morphology variations and complex defect modes.
• Quantitative Defect Scoring: Prototype algorithms segment and quantify defect regions in cross-polarized light (cracks, misoriented crystallites, high-strain fields), providing rapid macro-screening before downstream processing.
• Autoclave Time-Series Mining: In-situ temperature logs and actuator control signals are correlated with final crystal quality to uncover micro-transients and thermal fluctuations undetectable from nominal recipes alone.

Current Progress

• Completed an in-depth audit of historical ammonothermal growth logs (Ammono database) and formulated hardware specifications for local GPU computing nodes and storage servers.
• Deployed first-generation Python/OpenCV prototypes for automated crystal boundary segmentation and defect area fraction calculation.
• Established standardized schemas for high-frequency autoclave time-series logs and structural defect taxonomy in close collaboration with WP1 process engineers.