LAUSR

Abstract Numerical modelling is considered a standard tool for problem-solving in modern engineering. Based on various mechanisms, researchers have developed a collection of numerical models to simulate a complex system’s responses and provide forecasts for future physical conditions. However, deviations between the performance of each single model and real-world observations always exist. A man-machine model (MMM) is proposed for more accurate predictions based on multi-model ensembles. In the MMM, the man part reflects the usage of human-designed numerical models, and the machine part represents the application of machine learning algorithms. The MMM is depicted by a three-layered framework: an input layer loaded with simulations and target observations, a hidden layer composed of optimal factors, and an output layer that delivers approximations and forecasts. First, exploratory factor analysis is utilized to extract candidate factors from a variety of numerical models. Subsequently, an adjusted “General to Specific” rule is applied to select optimal factors in the process of best fitting the principal components of observations. Thereafter, the MMM is used to deliver a pattern recognition algorithm involving a Linear-Gaussian kernel function projecting the hidden layer to the observations. A case study of a beam loading test shows that the MMM is successful in giving robust predictions of deformation on testing data and avoids overfitting.