This work presents a deep learning-based approach to tackle inverse heat conduction problems in one and two dimensions from the available temperature profile and heat flux. The primary objective is to predict the equivalent thermal conductivity of materials in different compositions. Three distinct cases are presented: a uniform thermal conductivity material, a checkerboard pattern with alternating conductivity, and a composite structure comprising dispersed phase and matrix with varying conductivity. A comprehensive noise analysis is conducted on the one-dimensional heat conduction case, wherein input data is subjected to 5%, 10% and 25% noise levels. This analysis explores the model’s robustness and sensitivity to noisy input.