Domain-informed Machine Learning for Smart Manufacturing

Product geometric quality is critical to product functionality and proper fit. In AM, the post-production measurement of product accuracy is often conducted through 3D scanning using structured light scanning or laser scanning. A cloud of points will be generated by the 3D scanner to digitally replicate the scanned object or manufactured part. Embedded in the Euclidean space, each point has its set of Cartesian coordinates measuring a point on the exterior surface or boundary of an object. Depending on the sensing and scanning techniques, each point may also contain RGB color data or intensity information. Point clouds have been widely used in many engineering and medical disciplines such as computer-aided design, reverse engineering, metrology and quality inspection, architecture modeling, geographic information systems, augmented reality, and medical imaging.

Learning and prediction of geometric quality in AM aims to learn and predict geometric shape deviations of both built and untried product shapes based on a small set of training products. To effectively control shape deviations of new and untried product shapes in AM, we classify the modeling approaches as predicting modeling and prescriptive modeling. While traditional predictive modeling usual makes prediction within its experimental domains, e.g., a class or family of products, prescriptive modeling is able to make prediction of quality of new and untried categories of shapes beyond the experimental scope. This chapter will first introduce predictive modeling approach to predict geometric quality for products with simply 2D geometries. The modeling approach is further extended and generalized for prescriptive small-sample modeling and prediction for 2D freeform products.

The final product geometries in AM are often deformed or distorted. The deviations of three-dimensional (3D) shapes from their intended designs can be represented as 2D surfaces in a \(\mathbb {R}^3\) space, which constitutes a complicated set of data for representing, learning, and predicting geometric quality. Patterns of deviation surfaces vary with shape geometries, sizes/volumes, materials, and AM processes.

Three-dimensional (3D) shape accuracy is a critical performance measure for products built via AM. With advances in computing and increased accessibility of AM product data, machine learning for AM (ML4AM) has become a viable strategy for enhancing AM. Proper description of the 3D shape formation through the layer-by-layer fabrication process is critical to incorporate process domain knowledge into ML4AM. The physics-based modeling and simulation approaches present voxel-level description of an object formation from points to lines, lines to surfaces, and surfaces to 3D shapes. However, this computationally intensive modeling framework does not provide a clear structure for machine learning of AM data. Chapter 5 introduces domain-informed small-sample learning and prediction of shape accuracy of 3D objects.

While predicting shape deviation belongs to the forward problem, minimizing shape deviation of AM-built products is a challenging inverse problem due to geometric complexity, product varieties, material phase changing and shrinkage, interlayer bonding, and limited sample data. Various methods and strategies have been developed to improve geometric quality of AM processes. We summarize the related work in Table 6.1 with a sample literature given for each category.

This chapter introduces applications of the optimal compensation method and algorithm to reduce the shape deviation in metal AM process and to statistically monitor the part-to-part changes in AM.

Recent advances in the industrial Internet of Things and Cyber–Physical Systems have resulted in greater connections and accessibility of smart manufacturing. A particularly exciting consequence is the development of a new paradigm of smart AM systems that seamlessly integrate computing, manufacturing, and services. Each individual AM or 3D printing machine in such a system enables direct manufacturing of complex shapes from CAD models with reduced labor and costs compared to traditional manufacturing methods. The impacts of such systems are not yet fully realized in practice because their constituent processes may yield inconsistent product quality. Furthermore, Individual processes and machines face varying degrees of insufficient data and physical knowledge. A specified process model through machine learning often has a limited scope of application across the vast spectrum of processes in a manufacturing system that are characterized by different settings of process variables, including lurking variables. Knowledge or model transfer among AM processes is therefore essential to smart AM systems.

A significant challenge in quality control of an AM system is the model specification for different computer-aided design products manufactured by constituent AM processes. Current machine learning techniques can require substantial user inputs and efforts to implement in practice.

Nanomanufacturing represents the future of US manufacturing. Nanostructured materials and processes have been estimated to increase their market impact to about $340 billion per year in the next 10 years. In the past decade, tremendous efforts have been devoted to basic nanoscience discovery, novel process development, and concept proof of nano devices. Yet much less research activities have been undertaken in nanomanufacturing to duplicate the success of transforming quality and productivity performance of traditional manufacturing. High cost of processing has been a major barrier of transferring the fast-developing nanotechnology from laboratories to industry applications. The process yield of current nano devices is typically 10% or less. Hence, there is an imperative need of process improvement methodologies for nanomanufacturing.

This chapter introduces domain-informed modeling of nanowire (NW) growth process at multiple scales of interest for prediction. The main idea is to integrate available data and physical knowledge through a Bayesian hierarchical framework with consideration of scale effects. At each scale, the NW growth model describes the time-space evolution of NWs across different sites on a substrate. The model consists of two major components: NW morphology and local variability. The morphology component represents the overall trend characterized by growth kinetics. The area-specific variability is less understood in nanophysics due to complex interactions among neighboring NWs. The local variability is therefore modeled by an intrinsic Gaussian Markov random field (IGMRF) so as to separate itself from the growth kinetics in the morphology component. Case studies are provided to illustrate the NW growth process model at the coarse and fine scales, respectively.

Understanding the nanostructure growth faces issues of limited data, lack of physical knowledge, and large process uncertainties. These issues result in modeling difficulty because a large pool of candidate models almost fit the data equally well. Through the domain-informed machine learning strategy, we derive the process models from physical and statistical domains, respectively, and reinforce the understanding of growth processes by identifying the common model structure across two domains. This cross-domain model building strategy essentially validates models by domain knowledge rather than by (unavailable) data. It not only increases modeling confidence under large uncertainties, but also enables insightful physical understanding of the growth kinetics. We present this method by studying the weight growth kinetics of silica nanowire under two temperature conditions. The derived nanowire growth model is able to provide physical insights for prediction and control under uncertainties.

As a zero-band semiconductor, graphene is an attractive material for a wide variety of applications such as optoelectronics. Among various techniques developed for graphene synthesis, chemical vapor deposition (CVD) on copper foils shows high potential for producing few-layer and large-area graphene. Since fabrication of high-quality graphene sheets requires the understanding of growth mechanisms, and methods of characterization and control of grain size of graphene flakes, analytical modeling of graphene growth process is therefore essential to controlled fabrication. The graphene growth process starts with randomly nucleated islands that gradually develop into complex shapes, grow in size, and eventually connect together to cover the copper foil. To model this complex process, this chapter introduces a physical–statistical approach under the assumption of self-similarity during graphene growth. The growth kinetics is uncovered by separating island shapes from area growth rate.

Graphene is an emerging nanomaterial for a wide variety of novel applications. Controlled synthesis of high-quality graphene sheets requires analytical understanding of graphene growth kinetics. The graphene growth via chemical vapor deposition starts with randomly nucleated islands that gradually develop into complex shapes, grow in size, and eventually connect together to form a graphene sheet. Models proposed for this stochastic process do not, in general, permit assessment of uncertainty. This chapter introduces a stochastic modeling framework for the growth process and Bayesian inferential models. The modeling approach accounts for the data collection mechanism and allows for uncertainty analyses, for learning about the kinetics from experimental data. Furthermore, we link the growth kinetics with controllable experimental factors, thus providing a framework for statistical design and analysis of future experiments.

Since properties of nanomaterials are determined by their structures, characterizing nanostructure feature variability and diagnosing structure defects are of great importance for quality control in scale-up nanomanufacturing. It is known that nanostructure interactions such as competing for source materials during growth contribute strongly to nanostructure uniformity and defect formation. However, there is a lack of rigorous formulation to describe nanostructure interactions and their effects on nanostructure variability. This chapter introduces a method to relate local nanostructure variability (quality measure) to nanostructure interactions under the framework of Gaussian Markov random field. With the developed modeling and estimation approaches, we are able to extract nanostructure interactions for any local region with or without defects based on its feature measurement. The established connection between nanostructure variability and interactions not only provides a metric for assessing nanostructure quality but also enables a method to automatically detect defects and identify their patterns based on the underlying interaction patterns. Both simulation and real case studies are conducted to demonstrate the developed methods. The insights obtained from real case study agree with physical understanding.

Nanostructure interactions contribute strongly to structure uniformity and defect formation. Characterizing the interaction not only assists to understand the growth kinetics but also provides a metric to benchmark growth quality. Chapter 15 presented interaction modeling and estimation methods for one local region, which assume complete feature measurement. In order to map interaction patterns across the whole substrate, direct application of the methods would require complete information of a specimen. This implies a formidable metrology task using current inspection techniques such as scanning electron microscopy (SEM). In this chapter, we relax the metrology constraint and analyze nanostructure interactions with incomplete feature measurement. Specifically, we optimize Expectation-Maximization (EM) algorithm based on Markovian properties of interaction modeling and develop a tailored space filling design to select which sites to measure.