An artificial Intelligence approach in the mechanical and morphological analysis of silica aerogels

Kurzfassung

Silica aerogels can be computationally modelled by means of the 3D diffusion-limited cluster-cluster aggregation (DLCA) algorithm. The structural and mechanical properties of the computationally modelled silica aerogels can be characterised by analysing the simulated models and performing finite element (FE) calculations. However, the modelling process of DLCA is “slow” and together with the FE simulations requires significant computational time. The objective of this thesis is to introduce a machine learning approach to eliminate the necessity of generating and simulating 3-d silica aerogel models for predicting their structural and mechanical properties. To this end, an artificial neural network (ANN) based on supervised learning is developed to perform a mapping between the DLCA model parameters, the aggregate's fractal dimension and it's stress-strain response. In the next step, the ANN is trained by data-sets consisting of 3-d network structures of silica aerogels, that are modelled by means of the DLCA algorithm with varying structural parameters, and their FE analyses. The trained ANN in combination with a reinforcement learning (RL) agent is then used to perform data driven structure property relationships on the silica aerogel aggregates. It was found that a artificial neural network can effectively map the DLCA input variables to the target material properties. The ANN had a R2 score of 0.841. For the inverse design of the aerogel micro-structure Reinforcement Learning as an optimisation problem was utilised. A soft actor critic agent was utilised to optimise the DLCA parameters so as to achieve the desired material properties of elastic modulus and fractal dimension. The RL agent was proven to be effective in predicting the DLCA input parameters and as compared to previous work done on the inversion of the neural networks, was also not dependent on the prior knowledge of the desired input space so as to constraint it.