Periodic Reporting for period 1 - ALIAS (Machine Learning for Structural Integrity Assessments)
Periodo di rendicontazione: 2021-07-01 al 2023-06-30
The continued safe operation of critical infrastructure is key in ensuring economic prosperity. The drive towards carbon neutrality presents significant challenges to engineers as increasing operational efficiency often results in harsh, unfavourable operating conditions. The cost of implementing more efficient processes is a reduction in material performance resulting in prohibitively short component lifetimes. Material damage is multifactorial and depends strongly on the material chemistry, loading conditions, and operating environment. Recent advances in machine learning offer a potential solution to uncovering the complex interactions that inform the material damage process. By bridging the knowledge gaps currently impeding SIA improvement engineers can safely operate key infrastructure more efficiently and continue operating for longer periods of time.
In this research we created a framework that automates the selection of ductile damage model parameters from a simple engineering test. The framework uses Bayesian Optimisation, a machine learning technique, to find a global solution that minimises the difference between a finite element simulation and experimental data. The Bayesian Optimisation technique uses a small initial dataset to generate a statistically based understanding of how simulated material response relates to a combination of ductile damage parameters. The overall objective of the framework is to minimise differences between the simulated material response and the experimental test. Based on the statistical inference a new combination of ductile damage parameters is selected, simulated and the material response is re-evaluated. As the number of simulations increase so to does the machines knowledge about the relationship between error and parameters. This enables the machine to ‘learn’ which parameter combinations will minimise the error between simulated material response and experimental data. Our framework has been successfully deployed, providing parameter combinations that have less than 2 % error between simulated response and experimental data. This research has been disseminated at two international conferences. A journal publication of this work is currently underway.
This research demonstrates the ability of machine learning tools to enhance our fundemental understanding of material damage. Here we have created a framework that enables engineers to autonomously select ductile damage parameters that simulate the material response to accurately represent the response of an experimental test. Once initiated the framework, which can operate on a local machine (i.e. laptop) as a background task. This enables engineers to focus on tasks that require creativity and intelligence.