Domain adversarial neural network to determine unbiased Bingham parameters

Abstract

The workability of cement-based materials should be scientifically analyzed through the rheological behavior. The Bingham equation is generally adopted to describe the flow of cement-based materials. Rheometer is a device that measures the two parameters of the Bingham equation. Numerous rheometers are used with various geometries and measuring protocols. However, the rheological measurement provides different Bingham parameters for a single material when rheometers, geometry or even the measuring protocol is replaced. This study aims to construct a model that can yield the ideal Bingham parameters, which are identical for a single material regardless of the measurement conditions, with inputs of raw measurements. This study first introduces the generation of ideal domain using the theoretical Bingham model. Then, the unsupervised domain adaptation allows the model predicting ideal Bingham parameters with inputs of raw measurements. The experiments were conducted with mortar mixtures measured by three different measuring protocols. The proposed method significantly reduced the Bingham parameter discrepancy. We also investigated the effects of the measurements number on the resultant models. Owing to the ideal domain generation and physical properties of variables, the adversarial training was also successfully conducted using 10 measurements representing the distribution of the corresponding measurement domain. Finally, the proposed method was applied to the concrete dataset, 12 mixtures, expected to have broad range of Bingham parameters, measured by BML, BTRHEOM, IBB rheometers. The proposed method provided nearly identical values regardless of the used rheometer.