We live in an age of unprecedented technological growth. This growth has changed the daily lives of people across the globe. Changes like more compact devices, improved battery life, and faster processing are accompanied by additional methods of communication and access to/sharing of information. The former are underpinned by exploiting new materials and advanced manufacturing technologies while the latter come at the hand of increasingly sophisticated artificial intelligence (AI). Although there have been incremental advances in leveraging AI for specific analysis tasks in advanced manufacturing, there are no established, generalizable frameworks for accelerating research and development across material systems and manufacturing processes.

Because of the physical regimes in which they operate, advanced manufacturing processes are supported by nascent first-principle simulation capabilities instead of the more conventional or established approaches due to their cutting-edge nature. Owing to the ongoing emergence of such simulation or physics-based models for advanced manufacturing processes, practitioners are often forced to rely on their intuition and a small body of data when designing experiments which may or may not probe synthesis regimes that ultimately result in desires material microstructures and performance metrics. This can lead to a slower, more expensive research and development cycle and a delay of advanced manufacturing deployment for pilot and commercial-sale applications.

It is well known in the materials science and manufacturing fields that material microstructures play a central role in associating manufacturing process parameters used in synthesizing a component (or sample) and its performance. As such, microstructural features are crucial for guiding and interpreting manufacturing data. Of the multiple methods which enable determination of metal microstructures, scanning electron microscope (SEM) imaging is a popular approach for capturing information regarding important material features such as grain size distribution, precipitate morphology, and grain boundary density amongst others. However, SEM images must be analyzed to identify key features of interest, which requires domain knowledge and post-processing activities. All this makes SEM imaging a time and resource-intensive endeavor. Accordingly, there is great interest in reducing the number of SEM images that have to be obtained for a developmental process while also decreasing the cost of associated post-processing and analysis efforts.

There are several models available for predicting the microstructures of materials manufactured in a specific process parameter regime or identify the microstructures of the materials demonstrating a specific combination of performance metrics using first-principles approaches for conventional manufacturing approaches. However, as discussed above, such models are readily available to generate microstructures corresponding to either specific process conditions or final performance for advanced manufacturing. More recently, deep learning (DL) has found various applications to interpreting and understanding SEM images in the materials science and manufacturing applications, such as automatic classification of images azimi2018advanced ; muller2020bainitic ; 2020steel and segmentation of images to identify different regions of interest durmaz2021inference . Recently, DL methods have been used to generate SEM images of different materials iyer2019conditional and more baskaran2021adoption ; however, it is important to note that in most of these works, the DL approach deals with images in isolation and is not explicitly informed by the manufacturing technique. It is well understood that microstructural features strongly depend on the manufacturing conditions used to produce them. Therefore, while it is a great advancement to use DL for generating SEM images, reducing the cost of associated research and development activities, we also note that these prior works are unable to incorporate the valuable, process-dependent information necessary to expedite the development-validation cycle. Subsequently, there is a critical need to generate SEM images conditioned on specific manufacturing parameters as the next wave of DL development for materials and manufacturing image analysis. Incorporating a conditional component into SEM image generation enables the production of synthetic SEM imagery which conditionally depends on either manufacturing process parameters or target material properties as illustrated in Figure 1. While conditional image generation models have been widely used, most DL techniques are data-hungry, which presents problems when applied to domains like materials science and advance manufacturing which suffer precisely from scarce data. Therefore, it is essential for any conditional SEM image generation to be feasible even when trained on small datasets.

In this paper, we present a rationale and approach for addressing these challenges in applying DL for SEM imagery given limited training data. We demonstrate an ability to produce realistic microstructures in synthetic images and provide methods to quantify consistency with experimental SEM images through the use of topological feature extraction. This work takes a critical step towards leveraging machine learning to help accelerate advanced manufacturing research and development in light of developing first-principles simulations. We develop generative models trained on SEM images of aluminum alloy AA7075 tubes manufactured via the Shear-Assisted Processing and Extrusion (ShAPE) technology WHALEN2021699 ; shaped1 . ShAPE is an emerging advanced manufacturing process that an synthesize rods, bars, tubes, and wires kalsar2022microstructureC ; li2022manufactureA ; li2021copperD ; reza2022effectB of different cross-sectional areas and shapes from metallic (pure metals, alloys) feedstock in various forms such as powders, chips, films, discs, and solid billets darsell2018shearG ; taysom2022fabricationE ; wang2020microstructuralF . ShAPE-synthesized parts demonstrate unique microstructures with minimal porosity and never-before-seen performance. Several publications are available describing the synthesis and characterization of ShAPE samples made from aluminum, magnesium, copper, and steel, among others jiang2017frictionI ; whalen2019magnesiumH . ShAPE demonstrated enhanced performance in bulk-scale components, making their scale-up pathways relatively viable for industry. Therefore, there is an urgent need to develop models which can associate ShAPE process parameters with resulting microstructures in order to reduce research and development time and deployment delays for ShAPE at an industrial scale.