Abstract Scope |
Structural materials in nuclear energy applications are often subjected to simultaneous evolutions in compositional, structural, and environmental condition spaces, imposing significant challenges on mechanical behaviors modeling. In this work, a new artificial neural network architecture, consisting of temporal convolutional neural network and fully connected neural network (TCN-FCN), was applied for learning complex stress-strain data. The causal convolution operation implemented in this TCN-FCN architecture can correlate the most informative loading history information to current stress state in a high-dimensional material parameter space. The TCN-FCN models were benchmarked with similar gated recurrent unit (GRU)-based recurrent neural network models and showed ~50% error reduction in modeling complex loading histories and high-dimensional dependencies (e.g., temperature, strain rate, materials conditions, etc.). Such TCN-FCN architecture demonstrates excellent generalization ability and universal capability in modeling high-dimensional complex stress-stress data, thus offering a robust alternative to conventional empirical/semi-empirical models for nuclear materials modeling and optimizations. |