Conference Agenda

Metal additive manufacturing (MAM) has received significant attention in recent years due to its significant advantages such as increased design flexibility for complex geometries, shorter production-cycle, and efficient use of raw materials. To fully realize the potential of MAM in the context of Industry 4.0, it is necessary to address challenges related to the mechanical reliability of printed parts and their associated costs. Currently, trial-and-error methods are the most common way of optimizing MAM process conditions for achieving the desired printing quality. Meanwhile, numerical simulations can provide a more profound understanding of the physical phenomena involved in the build process, leading to a more systematic optimization of process conditions, and ultimately making the `first-time-right' high-quality production possible. Achieving a thorough quantitative understanding of the process requires insights from models covering various physical aspects including thermal, mechanical, metallurgical, and fluid-dynamics interactions. However, high-fidelity simulations of such models are accompanied by significant computational costs and therefore have limits in applications, particularly in sensitivity and optimization analyses where solutions for a wide range of scenarios are required.

To address this challenge, we initiated a project in 2021, with the support of the Swiss National Science Foundation (SNSF), to explore the feasibility of meaningful acceleration of these simulations without significant compromise in accuracy and reliability. Specifically, the project aims to develop solutions for thermal, microstructure, and residual stress simulations for the laser powder bed fusion (LPBF) process. To generate experimental validation data, Hastelloy X serve as the 'model material'. An overview of the results obtained so far, focusing on thermal and microstructure simulations, are presented.

Several techniques have been examined to reduce the computational cost of thermal simulation for LPBF, including a multi-scale simulation strategy, surrogate modelling, and physics-informed neural networks (PINNs), where the advantages and limitations of each approach are discussed. In the field of microstructure modelling, a 'Neural Cellular Automata' method has been developed, which outperforms the conventional Cellular Automata with up to 6 orders of magnitude acceleration in computation speed. Moving forward, the project will continue with a focus on the development of affordable high-fidelity models of residual stress development until 2025.

Additive Manufacturing (AM) and especially its metal variants constitute today a reality for the fabrication of high-performance industrial components. In particular, AM allows the construction of novel cellular structures, the so-called lattices, where well-designed unit cells are periodically repeated over a macro-shape to achieve exceptional specific performances, such as unprecedent stiffness-to-weight ratios. These structures, however, are very difficult to simulate numerically: on the one hand, the application of multiscale methods based on homogenization appears delicate due to an insufficient separation of scales (macro versus cell scales); on the other hand, solving directly the high-fidelity, fine-scale problem requires handling large numbers of complex cells which is often intractable if standard methods are blindly used. As a solution, immersed domain techniques have been applied, but such methods, generic in terms of applications, may not be optimal in the case of lattices.

In this context, the purpose of this work is to develop a HPC algorithm dedicated to lattices that takes advantage of the geometric proximity of the different cells in the numerical solution. In order to do so, we start by adopting the CAD paradigm based on spline composition along with its corresponding IGA framework. This offers (i) great flexibility to design any lattice geometry and (ii) fast multiscale assembly of the IGA system. Then, we resort to the family of Domain Decomposition solvers, and develop an inexact FETI based algorithm that avoids solving numerous local cell-wise systems. More precisely, we extract the “principal” local cell stiffnesses using a greedy approach, and use the latter as a reduced basis to efficiently solve all the cell-wise systems. It results in a scalable algorithm that tends to be matrix-free. During the talk, a range of numerical examples in 2D and 3D will be presented to account for the efficiency of our method both in terms of memory and computational cost reduction.

Discontinuous fiber reinforced composites are used in many application areas ranging from automotive to healthcare. Such parts are often manufactured in and injection molding process, as it is an economical process for high volume markets. The simulation of the injection molding process is well established and specific commercial tools have been developed for this task. However, the transient solution of the underlying non-linear multi-phase flow is computationally expensive and computation may take multiple hours for complex geometries. This computational time is prohibitively large for computational optimization of the product design or the process parameters. Hence, we propose a two-step process to accelerate the mold filling prediction: i) Solve a modified Eikonal equation to compute distance maps to the injection gate and nearest walls. This is computationally cheap, as it is only a stationary equation to solve. ii) Train feed forward neural networks to obtain a data-driven relation between the encoded distance maps and mold filling features, such as fill time and fiber orientation. We sample a set of geometries, automatically generate CAD models, and simulate these in a commercial injection molding solver to build a training data set. Subsequently, we apply different feed forward neural network architectures and evaluate their performance.