Weigh-in-motion (WIM) accuracy in measuring static axle loads is affected by vehicle dynamics and noise. Neural networks can identify underlying relationships, such as the spatial repeatability in axle dynamics, and can efficiently remove noise. Furthermore, they can adapt to changing circumstances (e.g., traffic characteristics, road profile, or sensor failure), unlike conventional WIM calibration algorithms. The paper performance of a multilayer feed-forward artificial neural network algorithm applied to a multiple-sensor WIM is analyzed. Numerical simulations of the axle forces applied on a smooth road profile are used to train, validate, and test the artificial neural network algorithm. This dynamic axle load variation is predicted with the vehicle simulation model VESYM. The mechanical parameters of the truck models and their speeds are randomly varied over a range established from real traffic measurements. Once the theoretical WIM data are obtained at the sensor locations, the measurements are artificially corrupted with noise up to the typical level of WIM accuracy. Details are given on the process of the neural network design, the size of the training sample, and the length of the training period. The artificial neural networks approach resulted in higher accuracy than the traditional average-based calibration method, especially at high noise levels. As a result, it shows promise for estimating static axle loads from multiple WIM measurements.