Data-driven modeling of thermal history in additive manufacturing

Abstract Additive manufacturing (AM) has the potential to construct complex geometries through the simple and highly repetitive process of layer-by-layer deposition. The process is repetitive and fully automated, but the interactions between layers during deposition are tightly coupled. To unravel these interactions, the computational models of the manufacturing process are critically needed. However, current state-of-the-art physics-based models are computationally demanding and cannot be used for any realistic optimization. To address this challenge, we built a surrogate model (SM) of thermal profiles that significantly reduced the computational cost. We built this model based on the observation that any AM process exhibits a high level of redundancy and periodicity, making it an ideal problem for machine learning and surrogate modeling. We introduced a unique geometry representation that is the key insight for this work. Rather than directly using the part geometry, we directly use the gcode and translate it into a set of features (local distances from heat sources, e.g. extruder, and sinks, e.g. cooling surfaces). This set of features is directly used as an input for the SM of thermal history. Since this set can be calculated a priori from GCode, the explicit geometry considerations are largely factored out. Moreover, we leveraged the analytical solution to the moving heat source model to determine heat influence zone (HIZ). The size of HIZ allows deciding a priori what should be the cardinality of the distance sets. We showed that for fused filament fabrication, the size of HIZ is small; thus, the number of input variables for the SM is small as well. To build the SM, we first generated the thermal data using a physics-based model and use it to train an artificial neural network model. We trained the SM and demonstrate its high predictive power and low computational cost. Specifically, we demonstrated the capabilities of our model to construct the thermal history for points at the interfaces between roads, with acceptable accuracy (error below 5.0%) in almost real time (0.034 s). With such performance, this model opens the possibility of optimization as well as process planning, and in-situ monitoring for closed-loop control.