-
3.1 Predictive modelling, simulation and performance assessment
Introduction
Predictive and realistic modelling is essential for the evaluation of in-service performance, quantification of damage, prediction of failure and remaining service life assessment (Subthemes 4.1, 4.2 & 4.3). Moreover, to address the need for effective operation, pre-emptive maintenance, and to allow the development of augmented machine learning and AI routines (Subtheme 3.2), accurate models of infrastructure are required to obtain predictive and informative numerical solutions.
Innovation
This project will develop advanced modelling techniques for infrastructure systems that enable more accurate results on the performance, health, and integrity of a system. It will develop rigorous constitutive material models which will enable estimation of real-time performance, adopting a multi-scale multi- physics modelling platform including nonlinear material and geometric analyses, large deformation, instability and deterioration including initiation and progression subject to varying initial and boundary value conditions. A host of constitutive models of interest to POs will be developed. For geomaterials, models will be developed for meta-stable, time-dependent behavior of porous media allowing for primary, secondary and tertiary creep and the simultaneous presence of air and water in the pore space. The gradual movement with time has been shown to a major impact on the progressive weakening and eventual failure of such materials. Recent examples are the failure in the Cadia tailings storage facility in Orange and Fundao tailings dam in Brazil. Other models developed will include an accurate representation of damage initiation, progression, arrest and coalescence using the phase-field technique as well as models for accurately estimating real-time performance structural elements subject to fire, blast and vibration, non-linear geometric analysis, buckling analysis and progressive deterioration analysis. This subtheme will also develop tools to quantify the effects of mechanical and environmental loading effects on pipes, correlate corrosion rates with pipe deformations and acoustic response by introducing electro-chemical phenomena to soil-pipe interaction models, develop advanced leak detection sensors and associated analytics to optimise the repaired and renewal of pipe networks following Leak Before Break concept.
Outcome
Accurate computer models of real complex infrastructure and geomaterial systems; effective tools for interactive real-time performance evaluation.
-
3.2: Physics informed artificial intelligence, machine learning and explanation
Introduction
Monitoring systems embedded in infrastructures generate data that contain information on usage, weaknesses, potential failure points, etc. However, these data streams often are so large that their analysis must be automated.
Deep learning techniques can find patterns in these data that can be used for prediction, understanding and decision making. We are particularly concerned with creating “explainable” models of data. That is, we wish the machine learning system not only to have high predictive accuracy, but to be able to explain why the prediction is made. In this way, engineers and policymakers will achieve a greater understanding of their infrastructure and maintenance requirements.Innovation
Deep learning methods have been highly successful in building models with high predictive accuracy but are essentially “black boxes”, whereas “computational discovery” methods based on physical laws build models in the form of field equations and qualitative simulations that are human-readable but rely on libraries of pre-built equation templates. This project will capitalise on the complementary work in Subtheme 3.1, coupled with the Bayesian reliability techniques (with Subtheme 4.2) to develop physics informed deep learning to allow predictive capacity and robustness beyond the domain of training data.
Outcome
Machine learning and AI tools for predictive diagnosis and analysis.
