ADVANCED MATERIALS & PROCESSES | MAY/JUNE 2024 16 annotations. RSI Pipeline Solutions and Pacific Gas and Electric Company are exploring the use of MicroNet to automatically evaluate replica microstructures from steel line pipe to meet new regulations. OCAS, a steel research center in Belgium, is predicting steel hardness using MicroNet encoder features directly and comparing the accuracy to measurements on features segmented by MicroNet segmentation models. ASM International is integrating MicroNet into an upcoming Microscopy Analysis Toolkit to automatically quantify any microstructure using the features extracted by the encoders. MicroNet models are proving to be a powerful, general purpose microstructure analysis tool that is seeing extensive growth for both research and industrial applications. ML automatic image analysis is increasingly being explored for additive manufacturing (AM) in situ process monitoring. AM has revolutionized manufacturing with its ability to create complex single-part geometries and enable rapid prototyping. However, quality control and certification has remained a challenge for AM because of high process variability and the large number of important process parameters. In situ monitoring promises to improve AM quality through early defect detection and by enabling adaptive process controls. Recently, a computationally efficient CNN capable of running in real-time was used to automatically detect laser-powder bed fusion (LPBF) printing defects such as delamination and splatter from a thermo- graphic camera during printing as illustrated in Fig. 2[4]. In another study, an LPBF machine was designed with in-situ sensors and cameras to monitor the build process and a CNN was used to analyze the images to assess part quality and validate specific internal geometries and build defects[5]. Another group developed a CNN to analyze layer-wise images during selective laser melting (SLM) of metal powders[6]. This enabled the on-line detection of defects and provided a basis for adaptive controls and quality assurance. These examples illustrate the potential of automatic image analysis in improving the efficiency and quality of and confidence in AM processes. MATERIALS DESIGN Machine learning models can connect materials processing and composition to structure and properties by finding empirical correlations in training data (e.g., steel composition and heat treatment history input and yield strength output). In recent years, such ML models have become powerful tools to accelerate materials design. First, despite a reputation for being “black box” models, ML PSP models can provide interpretable insights into the physical mechanisms of materials behavior (e.g., steel yield strength is correlated with carbon content, tempering temperature, and quench rate), which can provide design rules where physical models are not available. Second, these ML PSP models can make thousands of predictions in a fraction of a second. As a result, when applied to a search space (e.g., a range of compositions and heat treatment time and temperatures), millions of candidate materials can be screened in minutes. For these reasons, ML models are becoming essential components of modern materials design workflows. In 2023, Sasidhar et al. at the Max-Planck-Institut für Eisenforschung[7] in Germany utilized ML to identify trends in steel corrosion with composition and utilized the resulting design rules to develop a novel corrosion- resistant surface treated alloy for extreme environments. The authors began by using a large dataset of alloy corrosion metric measurements scraped from the literature[8] and trained an ML model for pitting potential, with composition and corrosive environment as inputs. They then ran an optimization routine to maximize pitting potential, starting from common steel, Ni, Al, high entropy alloy compositions, and observed the changes in composition (Fig. 3). The optimization process reproduced known composition effects on corrosion (i.e., the role of N or Mo in stainless steels and Mo in Ni-Cr alloys) while also producing new trends (i.e., large N or C contents in Ni-Cr alloys and transition metal high entropy alloys promote corrosion resistance). The authors then utilized this new insight alongside CALPHAD software to explore N or C supersaturation in the various alloy classes to serve as model alloy systems for further study. The key enablers for ML success in this work were the availability of ML-ready data to train the models and the alloy knowledge of the materials to complement the ML results with their years of experience and ICME tools. Fig. 2 — An efficient convolutional neural network detects defects by using a thermographic camera in real-time during laser powder bed fusion printing. Reprinted with permission from Ref. 4 under the terms of the Creative Commons CC BY license.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==