Employing Deep Neural Networks and High-Throughput Computing for the Recognition and Prediction of Vein-Like Structures

In this investigation, convolutional neural networks (CNNs) are leveraged to engineer a simple segmentation and recognition algorithm specialized for the delineation of complex, network-like morphologies-often termed "vein-like structures (VLSs)"-in scanning electron microscopy (SEM) imagery. These intricate formations frequently appear during the nitriding treatment of medium- to high-carbon alloy steels. To navigate the multifaceted characteristics of such architectures, CNN-based methodologies are synergized with high-throughput thermodynamic computations via Thermo-Calc. This integration aims to quantify both the theoretical upper bounds and the actual values of the VLSs. By establishing deep neural network models for both theoretical upper bounds and actual measurements, the gap between thermodynamics and thermokinetics in the nitriding process is bridged. Applying this amalgamated predictive schema to 8Cr4Mo4V steel, a groundbreaking departure from conventional paradigms that exclusively depend on thermodynamic calculation-based diffusion models is effectuated. The emergent model yields transformative implications for the metallurgical sector, paving the way for the refinement of future nitriding algorithms and enhancements in nitriding methodologies. A novel convolutional neural network algorithm is developed for segmenting and recognizing vein-like structures in scanning electron microscopy images of nitrided 8Cr4Mo4V steel. By integrating deep learning with high-throughput thermodynamic computations, this method links thermodynamics with thermokinetics, providing insights into the nitriding process and advancing metallurgical practices.image (c) 2024 WILEY-VCH GmbH