Transfer learning framework for fatigue life prediction of additively manufactured parts
Abstract
Additive manufacturing has the potential to make things more efficient because of its ability to build customizable products quickly and with fewer resources required. The issue that additive manufacturing faces is the inferior performance relative to conventionally manufactured parts in the long run, thus hindering their adoption in mission-critical applications. The main cause of this inferior performance in as-built parts is surface roughness, which can not be avoided due to process physics. The deep valleys in the surface act as stress concentration areas that become crack initiation sites, thus leading to part failure. The scarcity of fatigue testing data compounds this problem, as fatigue tests are expensive in terms of time and cost. Transfer learning is a machine learning technique used to handle the problem of smaller datasets for training by leveraging the model learned for a related task and using it for the target task. This work aims to use the transfer learning technique and fatigue test data to build a framework that learns from the surface roughness data and makes predictions about fatigue life. This framework can potentially solve the data scarcity problem in studying the fatigue behavior of additively manufactured parts, thus facilitating their adoption in a wide spectrum of applications with confidence.