Generative Adversarial Networks as a Method to Predict Stresses in Structures

Abstract

There have been continuous advances in the field of Finite Element Analysis (FEA) allowing designers, architects, engineers and the public at large increasing ease and access. The methods of implementation however, have remained largely unchanged since their inception in the 1960s relying predominantly on costly computational software and hardware to carry out time and computing resource intensive calculations. furthermore, alterations to any simulation inputs, constraints or design parameters potentially nullify the previous results and require subsequent additional simulations. While there have been strides made towards adaptive FEA software [such as Ansys Discovery live for instance] these too tend to be prohibitively more costly or resource intensive than their contemporary counterparts. As an additional consequence, the analysis of a component, structure or system is almost always done remotely, and ideally well before manufacture. Similarly, in situations which require active monitoring, the telemetry is required to be passed to remote systems capable of carrying out the FEA computations. With the advent and rapid development of Artificial Intelligence (AI), more specifically advancements in Artificial Neural Networks (ANNs) as a new toolset for the solution of complex problems, the question arises, "Can Neural Networks be trained to emulate Finite Element Analysis?". This is the basis on which this work centres: Utilising a conventionally generated FEA dataset, a conditional Generative Adversarial Network (cGAN) is "taught" the physical behaviour of platework of varying geometry and material properties and subjected to randomly placed varying loadings. The subsequent "trained" model is hence capable of generating predictions for arbitrary inputs which correspond to the domain of input on which it was trained. Three experiments resulted in separate cGAN generator models trained to infer deflections from forces, stresses from deflections and stresses from forces respectively. After a moderate training regime of 200 Epochs each, the outputs of the models are shown to be in reasonable agreement to the ground truth with mean errors in the range of 5-10 %. Whilst not perfect FEA replacements, the trained models show potential for improvement and in their existing implementation allow for near real-time iterations or testing of hypothetical force additions via a purpose built application. Furthermore, this adds credence to deploying systems which implement purpose trained models for the ablity to self monitor structures in situ in realtime.