Thermal interface materials (TIMs) are ubiquitous in electronic packages, wherein they are used to mitigate thermal contact resistances between solid-solid interfaces. TIMs are complex paste-like mixtures comprised of a base polymer solution in which dense metallic (or ceramic) filler particles are dispersed to improve the TIMs' heat transfer properties. Rheological characterization of such complex fluid-like materials is challenging due to their non-Newtonian flow behavior, which is not fully captured by classical shear-thinning models. We propose a data-driven approach using a physics-informed neural network (specifically, the rheology-informed neural networks (RhiNNs) proposed by Mahmoudabadbozchelou and Jalali (Sci. Rep., 2021). To demonstrate this approach, we first characterize the TIM as an elasto-visco-plastic (EVP) material, which has exhibits stress relaxation via an elastic shear modulus, has a yield stress, and has a constant shear viscosity upon flow. RhiNNs are used to solve the inverse problem of determining the EVP parameters to yield a dynamic response matching the experimental data within some tolerance. This training data is generated by  start-up flow experiments at different (constant) shear rates  using the TwinDrive MCR702 rheometer from Anton Paar.

In this way, we characterize the rheology of three different TIMs, each with significantly different observed flow behavior, including both flowability and elasticity. We validate the `learned' models by comparing their predicted shear stress evolution to experiments under shear rates not used in the training data sets. While only the EVP model was tested and benchmarked so far, the RhiNN approach enables rheological characterization of complex fluids with any general stress response input by the user, which is relevant for data-driven characterization of newer TIM formulations.