Research Projects

The project proposes to develop and validate stress-unsupervised machine learning frameworks to constitutively model heterogeneous materials and history dependent materials with high fidelity. These frameworks will only use surface displacement and boundary force data which are experimentally measured, unlike stress-tensor data which cannot be directly measured. Therefore, they will enable the widespread adoption of data-driven constitutive modelling in high-throughput experimental facilities, industries and biomedical devices. The first part of the proposal aims to develop and experimentally validate a machine learning framework to model heterogeneous hyperelastic materials. The proposed machine learning framework would involve simultaneous training of multiple neural networks, each modelling a heterogeneity in the material. The inputs to the neural network are the deformation gradients (strains), which can be experimentally measured, and the outputs of the neural network are strain energy densities. The neural networks are optimized by enforcing momentum equilibrium (stress-unsupervised). During initial development, this framework will use synthetic displacement and boundary force data from Finite Element simulations. Next, the machine learning framework will be used to determine the constitutive models of known materials having inclusions of known geometries and material properties. This will serve as an experimental benchmark for the developed framework. Experiments performed to this end would be quasistatic biaxial tension tests of elastomers (such as Rubber, Silicone, PDMS, etc.) with other elastomers as inclusions. Building on the previous framework, the next part of the proposal will develop and experimentally benchmark a machine learning framework to constitutively model history dependent material behavior- specifically viscoelastic materials. This framework would be built using Recurrent Neural architectures (Fourier Neural Operators and Recurrent Neural Operators) which would use experimentally available inputs (strains) and output the current stress tensor. These Recurrent Neural frameworks would be optimized by enforcing momentum equilibrium across multiple time-steps, making them more versatile compared to the previous framework. Since the Recurrent framework would be built around Neural operators, they will be invariant with the size of the time interval between strain inputs, unlike other existing frameworks. These frameworks will be benchmarked against diaphragm vibration experiments, wherein diaphragms made of elastomer supported hydrogels would be vibrated and the surface velocities are mapped using a laser doppler vibrometer by repeated tests. Successful implementations of these frameworks would enable real-time characterization of materials in additive manufacturing. It will also find use in detecting location of plaque and cancer tissue by rapidly measuring the local tissue stiffness and viscoelasticity.