Mediolateral ground reaction force(GRF) in walking can be used as an indicator of balance ability, prevention of injury, and diagnosis of disease and rehabilitation. Recently, as the need for wearable devices in everyday life, increase for the data utility with a small wearable device is required. On the other hand, methods for estimating the GRFs by simple measurement using a compliant walking model which can reproduce the GRFs of walking have been studied. Therefore, in this study, we propose the GRF estimation method using a compliant walking model that can reproduce 3D GRFs, and using machine learning based on the compliant walking model to improve the estimation accuracy. We estimated the unmeasured GRFs by selecting walking solutions which could reproduce a sacral marker position data. The GRFs were estimated using feedforward ANN model using sacral marker position data and model simulation. The NRMSE by using walking model were 22.1±8.0%(ML), 10.1±3.1%(AP), and 10.4±5.6%(vertical). The NRMSE with sacral positon data and the center of mass(CoM) position of model in the ANN model, were 11.9±3.4%(ML), 5.9±1.9%(AP), and 6.3±0.8%(vertical). However, there was no significant difference from sacral position data only. The result implies that the relationship of GRF and CoM motion of walking which could be represented by spring mechanics could help to select input of machine learning to estimate unmeasured walking data.