It is known that coda of strong motion records are products of numerous scatterings of body and surface waves within the subsurface soil structure. Several studies have successfully simulated coda wave envelopes by modelling the energy decay. However, due to limitations of quantifying soil heterogeneities, deterministic simulation of scattered wavefields is much more challenging. Therefore, the reverse problem of estimating non statistical properties of subsurface structure from coda waves remains a theoretical potential. On the other hand, machine learning techniques have proven useful in dealing with problems of similar nature, where a theoretical solution is imaginable yet hard to achieve due to a great number of unknown variables. This study utilizes artificial neural networks to propose a new approach of evaluating site effects from coda waves, with the future prospect of obtaining the similar results from microtremor records. A Long Short-Term Memory recurrent neural network is designed using Tensorflow 2 library in Python language. The study utilizes a strong motion dataset consisting of about 60000 3-component records as well as borehole data at 464 stations of Kiban-Kyoshin Network across Japan. The prediction input is coda wave timeseries of strong motion records, defined based on a parametric energy criterion, and all 3 seismograph components EW, NS and UD are used as parallel sequential features. In the first step, the prediction target is a vector of 3 site effect proxies namely, time-averaged shear-wave velocities for the upper 30-m depth Vs30 and the upper 10-m depth Vs10 and predominant frequency f0. In this step, different model parameter combinations are tested to ensure the basic model’s ability in extracting site-specific information from the input coda waves. One of the combinations is then used in the second step, in which the prediction target is surface to downhole ratio of Fourier Amplitude Spectra. For each of the 3 components EW, NS and UD, 100 identical networks are trained to each predict the desired ratio at a certain target frequency. Accuracy of test sample predictions confirms applicability of the proposed approach as well as its potential for future works on microtremor timeseries instead of coda waves.
