Architected materials are at the forefront of research as a way of producing high-performing structures with minimal weight. Moreover, the inclusion of nature-inspired hierarchical features has the potential to further enhance the performances of these materials. In this context, traditional honeycombs, extensively used in the aerospace industry, may benefit from the introduction of hierarchical levels. Additionally, higher-order features can be individually tailored to create a wide range of structures with peculiar mechanical responses. However, the design space becomes too large to be explored through conventional methods, necessitating the adoption of data-driven approaches. Thus, the present work aims to develop a methodology for generating 2D first-order hierarchical honeycombs with user-defined mechanical properties, leveraging the latest advancements in deep learning. The core of this procedure is the conception of an innovative generative model, extending Latent Diffusion Models (LDMs) to process lattice structures represented as graphs. While graph-based Latent Diffusion has been recently applied to the generation of synthetic or molecular graphs, this is the first instance of its application to the design of architected materials. Specifically, the structural graphs are encoded into a continuous latent space via a variational auto-encoder, thereby overcoming the complexity of discrete diffusion algorithms. Given the unique challenges of graph auto-encoding, the innovative Permutation Invariant Graph Variational Auto-Encoder (PIGVAE) is employed. Thereupon, a Denoising Diffusion Implicit Model, trained on latent graph embeddings, generates new 2D first-order hierarchical honeycombs. The generation is controlled through classifier-free guidance, which conditions the sampling process on the desired mechanical properties. Furthermore, the conditioning information is enriched by a combination of a contrastive-learning module and a diffusion-based Prior. The results demonstrate that the proposed methodology exhibits promising generative capabilities, and the model effectively captures underlying patterns in the original data.
I materiali architettati rappresentano una soluzione promettente per la realizzazione di strutture ad alte prestazioni e peso ridotto. Inoltre, l’integrazione di caratteristiche gerarchiche, ispirate alla natura, può migliorarne ulteriormente le prestazioni. In questo contesto, i nidi d’ape, ampiamente utilizzati nell’industria aerospaziale, potrebbero beneficiare significativamente dall’introduzione della multiscala. Le caratteristiche di ordine superiore, inoltre, possono essere personalizzate in modo mirato, ottenendo strutture meccanicamente uniche. Tuttavia, è chiaro che lo spazio di progettazione diventa troppo vasto per essere esplorato con strategie convenzionali, rendendo necessario l’impiego di approcci basati sui dati. Pertanto, questo lavoro si propone di sviluppare una metodologia per generare, in modo condizionato, nidi d’ape bidimensionali gerarchici di primo ordine, sfruttando le ultime innovazioni nel campo dell’apprendimento automatico al calcolatore. Il cuore di questa procedura è un modello generativo che estende i Latent Diffusion Model (LDM) al trattamento di strutture reticolari, rappresentate come grafi strutturali. Nonostante la diffusione latente dei grafi sia stata recentemente applicata alla generazione di grafi sintetici o molecolari, questo lavoro è la prima applicazione nel contesto della progettazione di strutture reticolari. In particolare, i grafi strutturali sono codificati in uno spazio latente e continuo, evitando le complessità degli algoritmi di diffusione discreta. A causa delle peculiarità dei grafi, si è scelto di utilizzare il Permutation Invariant Graph Variational Auto-Encoder (PIGVAE). Nuovi nidi d’ape gerarchici possono quindi essere generati da un Denoising Diffusion Implicit Model, addestrato sulla rappresentazione latente dei grafi. Questo processo generativo è condizionato sulle proprietà meccaniche desiderate, tramite classifier-free guidance, ed è arricchito da due moduli ausiliari. I risultati dimostrano che la metodologia proposta presenta interessanti capacità generative e che il modello è in grado di catturare relazioni profonde tra le strutture del dataset originale.
Deep conditional generation of hierarchical architected materials via graph latent diffusion
Pizzamiglio, Stefano Maria
2023/2024
Abstract
Architected materials are at the forefront of research as a way of producing high-performing structures with minimal weight. Moreover, the inclusion of nature-inspired hierarchical features has the potential to further enhance the performances of these materials. In this context, traditional honeycombs, extensively used in the aerospace industry, may benefit from the introduction of hierarchical levels. Additionally, higher-order features can be individually tailored to create a wide range of structures with peculiar mechanical responses. However, the design space becomes too large to be explored through conventional methods, necessitating the adoption of data-driven approaches. Thus, the present work aims to develop a methodology for generating 2D first-order hierarchical honeycombs with user-defined mechanical properties, leveraging the latest advancements in deep learning. The core of this procedure is the conception of an innovative generative model, extending Latent Diffusion Models (LDMs) to process lattice structures represented as graphs. While graph-based Latent Diffusion has been recently applied to the generation of synthetic or molecular graphs, this is the first instance of its application to the design of architected materials. Specifically, the structural graphs are encoded into a continuous latent space via a variational auto-encoder, thereby overcoming the complexity of discrete diffusion algorithms. Given the unique challenges of graph auto-encoding, the innovative Permutation Invariant Graph Variational Auto-Encoder (PIGVAE) is employed. Thereupon, a Denoising Diffusion Implicit Model, trained on latent graph embeddings, generates new 2D first-order hierarchical honeycombs. The generation is controlled through classifier-free guidance, which conditions the sampling process on the desired mechanical properties. Furthermore, the conditioning information is enriched by a combination of a contrastive-learning module and a diffusion-based Prior. The results demonstrate that the proposed methodology exhibits promising generative capabilities, and the model effectively captures underlying patterns in the original data.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Pizzamiglio_Tesi.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
27.1 MB
Formato
Adobe PDF
|
27.1 MB | Adobe PDF | Visualizza/Apri |
2024_12_Pizzamiglio_Executive Summary.pdf
non accessibile
Descrizione: Executive summary
Dimensione
1.07 MB
Formato
Adobe PDF
|
1.07 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/230481