Data-Defined Machine Intelligence for Mechanical Properties Identification

In this work we are introducing a Machine Intelligence (MI) concept that belongs to the set of approaches that emphasize data over artificial intelligence model. The development of the concept was motivated by the need to eliminate training as the means to enable fast or perpetual system adaptation and update, and also vastly reduce pre-processing times. This characteristic can be extremely valuable in situations when new data are regularly becoming available and utilized to increase the MI accuracy with minimal delay, or when the represented system changes with time. A characteristic of the proposed data-defined MI approach is that of introspection; it can be queried to locally estimate the system error at any data point that is part of its data set and potentially act accordingly. To provide evidence of the validity and performance of the approach, we first demonstrate on classical hand-writing classification of numerical digits. We then apply it to a challenging problem of identifying the properties of bi-linear elastoplastic material using a single full-field strain tensor image of a synthetic experiment. The validation test is performed for many untrained instances, and the histogram, mean and maximum errors are discussed. We also discuss several interesting capabilities that a data-defined concept enables.