This thesis presents a strategy based on artificial neural networks to compute optimal topologies for a 3-material Messerschmitt–Bolkow–Blohm (MBB) beam, starting from the design variables (force position and angle). The aim is to reduce the computation time with respect to a multi-material structural topology optimisation code for compliance minimisation, in particular, the PolyMat MATLAB code by [1]. The final neural network is built from the training of two independent models: a deep convolutional autoencoder and a fully-connected model for regression. The models were trained in the Google Colaboratory machine learning platform using the powerful TensorFlow-based Keras API. Further hyperparameter tuning was performed to improve the effectiveness of the fully-connected model, including a hyperparameter grid search and training with a step-decay learning rate. The obtained results show some typical issues arising with the implementation of machine learning models in topology optimisation, such as element disconnections and diffuse regions in some cases. Also, fine details such as the precise location of material interfaces might not be always predicted correctly. However, the predictions made by the neural network are sufficiently accurate to obtain a good estimate of the rough material distribution and contour of the final structural topology. Additionally, the computation time with respect to the reference optimisation code is significantly improved. Therefore, the proposed strategy could serve as an effective tool for obtaining optimal preliminary designs of a 3-material MBB beam within the limits of the design variables and the problem setting considered. Future developments are suggested in light of the limitations of the present work.
Questa tesi presenta una strategia basata su reti neurali artificiali per la predizione di topologie ottimali per una trave multi-materiale di tipo Messerschmitt–Bolkow–Blohm (MBB), a partire dalle variabili di progetto (posizione e angolo di applicazione della froza agente). L'obiettivo è ridurre il tempo di calcolo rispetto a un codice di ottimizzazione topologica multi-materiale, in particolare il codice MATLAB PolyMat di [1]. La rete neurale è ottenuta a partire dall'addestramento di due modelli indipendenti: un autoencoder convoluzionale e un modello completamente connesso per regressione. I modelli sono stati addestrati nella piattaforma di apprendimento automatico di Google Colaboratory, utilizzando la potente API Keras basata su TensorFlow. È stata eseguita un'ulteriore messa a punto degli iperparametri per migliorare l'efficacia del modello completamente connesso, inclusa una ricerca su griglia degli iperparametri e un addestramento con un tasso di apprendimento con decadimento graduale. I risultati ottenuti evidenziano alcuni problemi tipici dell'utilizzo di modelli di apprendimento automatico per l'ottimizzazione topologica, che in alcuni casi includono elementi disconnessi e regioni diffuse. Inoltre, dettagli fini come la posizione precisa delle interfacce dei materiali potrebbero non essere sempre riprodotti correttamente. Tuttavia, le topologie predette dalla rete neurale sono sufficientemente accurate per fornire una buona stima della distribuzione del materiale e della topologia strutturale ottimale. Inoltre, il tempo di calcolo rispetto al codice di ottimizzazione di riferimento è notevolmente migliorato. Pertanto, la strategia proposta può servire come uno strumento efficace per produrre topologie ottimali preliminari per una trave MBB a 3 materiali, entro i limiti delle variabili di progetto e dell'impostazione del problema considerati nel lavoro svolto. Infine, vengono suggeriti possibili sviluppi futuri alla luce dei limiti evidenziati nel presente lavoro.
multi-material topology optimisation exploiting artificial intelligence
CAMACHO CASTILLO, SANTIAGO
2021/2022
Abstract
This thesis presents a strategy based on artificial neural networks to compute optimal topologies for a 3-material Messerschmitt–Bolkow–Blohm (MBB) beam, starting from the design variables (force position and angle). The aim is to reduce the computation time with respect to a multi-material structural topology optimisation code for compliance minimisation, in particular, the PolyMat MATLAB code by [1]. The final neural network is built from the training of two independent models: a deep convolutional autoencoder and a fully-connected model for regression. The models were trained in the Google Colaboratory machine learning platform using the powerful TensorFlow-based Keras API. Further hyperparameter tuning was performed to improve the effectiveness of the fully-connected model, including a hyperparameter grid search and training with a step-decay learning rate. The obtained results show some typical issues arising with the implementation of machine learning models in topology optimisation, such as element disconnections and diffuse regions in some cases. Also, fine details such as the precise location of material interfaces might not be always predicted correctly. However, the predictions made by the neural network are sufficiently accurate to obtain a good estimate of the rough material distribution and contour of the final structural topology. Additionally, the computation time with respect to the reference optimisation code is significantly improved. Therefore, the proposed strategy could serve as an effective tool for obtaining optimal preliminary designs of a 3-material MBB beam within the limits of the design variables and the problem setting considered. Future developments are suggested in light of the limitations of the present work.File | Dimensione | Formato | |
---|---|---|---|
2023_05_Camacho.pdf
solo utenti autorizzati dal 18/04/2024
Descrizione: Thesis
Dimensione
5.15 MB
Formato
Adobe PDF
|
5.15 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/10589/203582