Speedup of FTMM (finite-size transfer matrix method) calculations through machine learning techniques

Abstract

Current methods to calculate the random incidence absorption of a finite size porous sample require complicated numerical integrations. An effort was made to replicate the FTMM (Finite-Size Transfer Matrix Method) using machine learning algorithms. A script was developed to generate 5000 normal incidence and FTMM solver files of 1 or 2 layer composites ranging from 5 to 100 mm on a commercially available solver. The FTMM sample size was 8'x9', as is used in the ASTM C423 standard. Various machine learning algorithms were then trained on half the dataset. The frequency and complex normal incidence impedance were used as inputs, and the FTMM C423 result was the output. The best result produced was a 7-layer neural network with a mean deviation of .03 from the full dataset after 1000 rounds of training, with significant faster computation compared to the commercial FTMM solver.