Page 40 - ESAM-1-1

Engineering Science in
Additive Manufacturing Gen-AI for lattice structure design

During the design process, the adjacency matrix and node validations, including FEM simulations and AM, confirmed
features were transformed through a forward diffusion that the generated structures closely aligned with target
process, and a transformer-based reverse denoising mechanical properties, underscoring the potential of
process was used to reconstruct graph structures that meet the GLU3D framework for efficient and accurate inverse
target design metrics such as stiffness and maximum stress. design in advanced manufacturing.
Comparative experiments demonstrated that GraphDGM
outperforms traditional methods in design accuracy, 3. Generative AI for performance
achieving superior results across multiple case studies. optimization of lattice structures
This framework offers a promising solution for efficient, Optimizing lattice structures for superior performance
multiobjective inverse design of complex structures. is a crucial objective in advanced manufacturing. The
With the increase of design requirements, inverse targeted properties optimized in lattice structure design
design is no longer limited to regular frame structures, are summarized in Table 1. Traditional design approaches
but has expanded to more complex and irregular TPMS often rely on iterative simulations and experimental
structures to improve performance diversity and optimize testing, which can be time-consuming and limited in
space. To address this, Li et al. proposed an inverse design design exploration. Recent advancements in Gen-AI,
52
framework for TPMS cellular structures by integrating a particularly GANs, have opened new possibilities for
bidirectional generative adversarial network (BiGAN) with performance-driven design optimization. By enabling the
a forward prediction model, termed the IOA-Model. The automatic generation of diverse, high-performance lattice
framework aims to generate combined TPMS structures structures, these approaches facilitate efficient design
(Primitive and IWP types) that meet specified mechanical space exploration and offer scalable solutions for tailoring
property targets, such as energy absorption and buffering mechanical properties.
capabilities. The BiGAN model maps structural parameters For instance, Challapalli et al. proposed an inverse
53
to a latent space and generates new structural configurations, machine learning framework utilizing GANs to design
while the physical model predicts mechanical properties lightweight lattice meta-materials for optimizing load
to guide the generative process. The designed structures carrying capacity, as shown in Figure 4. The framework
were fabricated using laser powder bed fusion (LPBF) and integrates GANs with forward regression models and
validated through FEM and compression experiments. The boundary conditions to generate novel lattice unit cells
results demonstrated that the generated structures closely with superior mechanical properties, specifically optimized
matched target mechanical behaviors, with minimal errors for compressive strength. The process begins by converting
in load-displacement curves and elastic modulus. This lattice structures into numerical fingerprints, which are
study highlights the potential of leveraging Gen-AI for then used for training the GANs and regression models.
efficient, high-precision inverse design of complex cellular The generated designs were validated through FEM and
structures. experimental compression tests using 3D-printed samples.
Although BiGAN-IOA has improved design complexity Theresults showed that the optimized lattice structures
and performance accuracy, challenges still exist in exhibited 40 – 120% improvement in compressive strength
implicit geometry processing, generation efficiency, and compared to traditional octet unit cells. Furthermore, these
comprehensive optimization. Jadhav et al. proposed unit cells were applied to design lattice-cored sandwich
35
the GLU3D framework to achieve more comprehensive structures, which also demonstrated superior mechanical
and efficient inverse design, which efficiently generates performance. The study highlights the framework’s ability
complex lattice unit cell structures based on specified for continuous optimization, offering a scalable approach
mechanical properties, as shown in Figure 3. The process for designing advanced load-bearing structures with
begins by generating implicit representations of lattice tailored mechanical properties.
structures, which are then seamlessly converted into mesh Similarly, Eren et al. proposed a deep learning-
54
structures for AM and voxel structures for FEM analysis. enabled design methodology for tailoring the mechanical
The framework incorporates a novel tanh-based scaling properties of metallic lattice structures fabricated via LPBF.
method to preserve geometric accuracy in the implicit They developed a 3D GAN (3DGAN) model trained on a
domain, and it introduces hybrid structures by combining dataset of lattice structures generated through parametric
various TPMS configurations. These generated structures design and optimized using a simulated annealing
demonstrated superior performance in energy absorption algorithm. The GAN model was employed to produce novel
and compression strength compared to conventional cubic lattice designs aimed at enhancing mechanical properties
structures, particularly at lower densities. Experimental such as stiffness, durability, and energy absorption. These

Volume 1 Issue 1 (2025) 5 doi: 10.36922/ESAM025110006