Page 43 - ESAM-1-1

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

critical slowing down. These limitations become especially 6. Summary and perspectives
pronounced when simulating intricate lattice geometries
under varying process parameters and loading conditions, 6.1. Summary
which require extensive iterations and high-fidelity Lattice structures, renowned for their exceptional strength-
analyses to achieve accurate results. to-weight ratios, energy absorption capabilities, and heat

Gen-AI offers a promising solution to these challenges dissipation properties, have found widespread applications
by accelerating simulations, reducing autocorrelation across aerospace, biomedical, and industrial sectors.
times, and enabling the generation of independent samples However, designing optimized lattices that meet complex
more efficiently. One effective approach involves the use performance criteria remains a significant challenge, often
of generative models, such as GANs and VAEs, to develop hindered by the limitations of traditional design methods
surrogate models that approximate the behavior of complex and computational inefficiencies. Overall, this review
lattice structures. These surrogate models can rapidly highlights the transformative role of Gen-AI in advancing
predict mechanical responses, thermal behavior, or stress the design and optimization of lattice structures for AM,
distributions, significantly reducing the need for resource- as presented in Figure 5. Gen-AI, particularly models such
intensive simulations. For instance, once trained on high- as GANs, VAEs, and diffusion models, has emerged as a
fidelity simulation datasets, these AI-driven models can powerful tool to revolutionize lattice design by automating
provide near-instantaneous predictions for new lattice inverse design processes, optimizing performance,
configurations, thereby expediting the overall design and considering process constraints, and accelerating
validation cycle. For instance, Pawlowski et al. proposed simulations.
57
a method to reduce autocorrelation times in simulations For inverse design, frameworks like GraphDGM and
51
by integrating GANs with traditional Hybrid Monte Carlo GLU3D demonstrate how Gen-AI can efficiently generate
35
(HMC) algorithms. In conventional HMC simulations, complex lattice configurations that meet predefined
critical slowing down often leads to long autocorrelation mechanical property targets. These models leverage
times, reducing the efficiency and scalability of the advanced neural architectures and diffusion models
simulation process. To overcome this, the authors to explore vast design spaces and propose optimized
introduced a GAN-based overrelaxation step to generate solutions that align with specific stiffness, stress, and
independent field configurations. This approach effectively energy absorption requirements.
breaks the Markov chain for observables unrelated to the In terms of performance optimization, Gen-AI models
action, allowing for faster convergence when maintaining enable the generation of lattice structures that significantly
statistical accuracy. The method was validated using the outperform traditional designs in mechanical strength,
two-dimensional scalar ϕ4-theory, where significant energy absorption, and strength-to-weight ratios. Studies
reductions in autocorrelation times were observed integrating GANs with simulation-driven methods and
without compromising ergodicity or simulation fidelity. experimental validations have showcased improvements
Comprehensive consistency checks further confirmed the of up to 120% in compressive strength and 150% in
robustness and accuracy of the GAN-integrated approach, elongation, demonstrating the scalability and efficiency of
demonstrating its potential to enhance computational AI-driven design processes. 54
efficiency in complex lattice simulations.
Process-aware lattice design has also been enhanced
In addition, the potential of Gen-AI in accelerating
FEM simulations is also noteworthy. Traditional FEM by Gen-AI models like ACGAN, which consider not
only the geometry of the lattice but also critical AM
simulations for lattice structures can be computationally process parameters (e.g., print speed, layer height, and
intensive, particularly when dealing with highly complex temperature). 56,58 Such integrated frameworks ensure that
geometries, multimaterial interfaces, or large datasets generated designs are not only structurally optimized but
involving multiple design variations. In such scenarios, also feasible for practical fabrication, leading to minimized
integrating Gen-AI models as surrogate predictors can deviations between predicted and actual performance
drastically reduce simulation times. For example, GANs
can be trained to learn the mapping between lattice outcomes.
geometries and their corresponding mechanical responses, For simulation acceleration, integrating Gen-AI
enabling the rapid prediction of stress-strain behavior, models with traditional FEM simulations has proven
deformation patterns, or failure modes. Similarly, diffusion effective in reducing autocorrelation times and enhancing
models can generate synthetic simulation results that computational efficiency. 57,59,60 This advancement addresses
closely approximate real data, further reducing the need critical challenges in simulating complex lattice behaviors,
for exhaustive simulations. particularly under stochastic conditions, thereby improving

Volume 1 Issue 1 (2025) 8 doi: 10.36922/ESAM025110006