MACHINE LEARNING-DRIVEN PREDICTION OF BUCKLING RESPONSE IN TUBULAR STRUCTURES VIA FINITE ELEMENT-GENERATED DATA

Abstract

This paper presents a machine-learning method for predicting buckling in tubular structures. A comprehensive set of force-time histories is generated within a design space characterized by highly nonlinear buckling behavior, using a calibrated finite element model. The neural network is fully connected, with key hyperparameters optimized before testing on an unknown dataset. This dataset includes measurements such as peak load, average load, and energy absorption error. The results show systematic rather than random simple error patterns that correspond with the physical process of structural collapse. To confirm this, finite-element simulations are performed with various geometric imperfections. These errors show similar properties, and the neural network's error characteristics are analyzed in detail through force-time curve comparison. The deviations from the neural network are found to have physical significance. Overall, the results suggest that this approach can reliably replicate crushing behavior, with sufficient accuracy to account for minor differences caused by small geometric defects.

Keywords : Investor-State Disputes, Treaty Modernization, Foreign Investment Regulation, Bilateral Investment Treaties, Sovereign Control, Arbitration Consent.