Modeling process–structure–property relationships in metal additive manufacturing: a review on physics-driven versus data-driven approaches

Abstract

Metal additive manufacturing (AM) presents advantages such as increased complexity for a lower part cost and part consolidation compared to traditional manufacturing. The multiscale, multiphase AM processes have been shown to produce parts with non-homogeneous microstructures, leading to variability in the mechanical properties based on complex process–structure–property (p-s-p) relationships. However, the wide range of processing parameters in additive machines presents a challenge in solely experimentally understanding these relationships and calls for the use of digital twins that allow to survey a larger set of parameters using physics-driven methods. Even though physics-driven methods advance the understanding of the p-s-p relationships, they still face challenges of high computing cost and the need for calibration of input parameters. Therefore, data-driven methods have emerged as a new paradigm in the exploration of the p-s-p relationships in metal AM. Data-driven methods are capable of predicting complex phenomena without the need for traditional calibration but also present drawbacks of lack of interpretability and complicated validation. This review article presents a collection of physics- and data-driven methods and examples of their application for understanding the linkages in the p-s-p relationships (in any of the links) in widely used metal AM techniques. The review also contains a discussion of the advantages and disadvantages of the use of each type of model, as well as a vision for the future role of both physics-driven and data-driven models in metal AM.

Data-driven Reference Trajectory Optimization for Precision Motion Systems

“…k=1 ⊆ R 2 is the input trajectory, f lm is the RK4 approximation of the linear model (9), with identified matrices C and D, T k is the transformation matrix between global and local coordinates in (10), µ k is the discretized maximum allowed deviation from the target geometry, W are the limits of the working space of the device, v max is the maximum allowable speed, and a max is the acceleration limit.…”

“…These design principles are motivated by the requirements of "Industry 4.0" where machines are no longer standalone setups and are instead connected in a network, making data more easily available [8]. This data can be exploited to improve performance; for example, it can be leveraged to create data-driven models, avoiding the labour intensive process of developing and maintaining first principles based ones [9]. Future manufacturing is also expected to be more distributed and personalized (e.g., "Manufacturing as a service") [10], [11], leading to smaller volumes for any given part, possibly produced across a pool of non-homogeneous machines.…”

Hardness Prediction of Laser Powder Bed Fusion Product Based on Melt Pool Radiation Intensity

“…Because the LPBF forming process is complex with several influencing factors, it is challenging to conduct research and analyses based on physical models. However, the data-driven approach has the advantage of modeling complex physical problems [10], and some classical machine learning models, such as support vector machines and linear regression algorithms, have been used for the classification and prediction of forming quality in additive manufacturing [11]. An important challenge of data-driven approaches for quality prediction is the accuracy and speed of prediction.…”

Investigating the Linear Thermal Expansion of Additively Manufactured Multi-Material Joining between Invar and Steel

“…Other studies have investigated the mechanical properties of SLM Invar and suggested process parameters to achieve a stable and refined microstructure with comparable mechanical properties to the traditionally fabricated alloy [ 22 ]. Powder-bed AM processes, like SLM, can produce high-resolution parts, but the build volumes of these systems are typically less than 0.3 m 3 [ 23 ]. Additionally, powder feedstock is very expensive, so these processes are not ideal for building large-volume components (such as molds and dies) [ 24 ].…”