Learning Based Parameter Estimation in Diffusion Models-An Identifiability Study

This work proposes a learning-based biosensing receiver (BSR) for estimating diffusion parameters in molecular communication (MC) systems. It is done from timely observations of concentration and first passage time signals (FPTD). CNN+LSTM and CNN+GRU sequence-to-vector models are used to estimate the diffusion coefficient and transmitter (TX)-BSR distance from analytically generated data. GRU achieves better performance, particularly for diffusion coefficient estimation. Results show that longer observation windows improve estimation accuracy because it captures both the early time exponential and late time scaling behavior, while parameter identifiability limits performance.