A Machine Learning Approach to Modelling Temperature-Dependent Cyclic Behaviour

This study presents a methodology for developing a temperature-dependent cyclic plasticity surrogate model as an efficient alternative to phenomenological temperature-dependent constitutive models. The titanium alloy Ti-6Al-4V, known for its widespread use in various engineering applications, was selected for this investigation. The surrogate model, based on a feedforward neural network, was trained using random amplitude stress-strain histories at various temperatures. To generate the training dataset, constitutive models were calibrated at specific temperatures using both experimental and available literature data, enabling the simulation of virtual temperature-dependent experiments. Cyclic loading simulations were performed at random axial strains within the range [-4%, 4%] and temperatures of 20℃, 400℃, 500℃, and 600℃. The predictive accuracy of the surrogate model was evaluated using unseen random stress-strain histories and temperature conditions, demonstrating high accuracy and computational efficiency.