Efficient Multiscale Multiphysics Modeling with Machine Learning Based Surrogate Models

In this study, a long short-term memory (LSTM) neural network was trained as a computationally efficient surrogate for a physics-based micromechanics model and embedded within a macroscale model based on classical lamination theory. Surrogate models are attractive because they can be evaluated many orders of magnitude faster than physics-based models and with a high degree of accuracy. Such models can be used for efficient multiscale modeling, design optimization, Monte Carlo methods, and optimal experimental design in ways that would be intractable with many physics-based models. The surrogate microscale model was trained from data generated using generalized method of cells to predict the homogenized stiffness of composite plies during loading with damage progression. The surrogate was able to learn a latent material damage representation to capture loading history and path dependence using the LSTM layers in the neural network architecture. When the surrogate was embedded within a physics-based macroscale model, the efficient model was able to make predictions 145 times faster than high fidelity generalized method of cells with a coefficient of determination of 0.98. A single LSTM surrogate model was trained to accurately predict the physical response of a composite ply with a wide range of material properties, but only for a fixed microstructure. Ongoing efforts to incorporate convolutional neural network (CNN) layers into the LSTM model in order to capture the effects of varying microstructure will be also be discussed.