Towards physics-consistent machine learning models: A geomechanics-based artificial neural network for high-cyclic soil response

Abstract

Estimating long-term soil deformation is crucial for the safe design of offshore foundations, railway substructures, pavement layers, and other cyclically loaded geotechnical systems. Traditional constitutive formulations, such as the High-Cycle Accumulation (HCA) model, provide reliable simulations but rely on empirically calibrated parameters that require extensive experimental campaigns for their determination. To overcome this shortcoming, it is proposed to design a Geomechanics-based Artificial Neural Network (GANN) that bypasses the need for calibration parameters and instead uses common soil descriptors. Two steps are required for this purpose: (i) develop a foundational GANN capable of accurately predicting accumulated strain evolution for a given soil; and (ii) generalize this GANN so that it can operate for a wide variety of soils using simple input variables such as the ones related to grain size distribution and state boundary surface. Accordingly, this manuscript aims to address this first step by considering two different soils as examples, namely Karlsruhe fine sand and Australian superfine silica sand. For each soil, a synthetic database is generated from the HCA model to train a GANN architecture, enabling it to learn the fundamental relationships governing strain accumulation. The methodology integrates data-driven learning with a physics-informed loss function, ensuring that the predicted strain evolution remains consistent with soil mechanics principles. This GANN comprises multiple fully connected layers using several activation functions, with outputs structured to capture strain accumulation over a high number of cycles. Validation against synthetic and experimental data confirms excellent performance, supporting this first step toward a generalized GANN framework.