SYNERGISING FINITE ELEMENT METHODS WITH AI-DRIVEN PREDICTIVE RISK MODELS IN MANUFACTURING: A SYSTEMATIC REVIEW
TL;DR
The evidence demonstrates that the benefits of hybridisation do not lie in replacing simulation with AI, but rather in ensuring predictions are more relevant and reliable using physics-based knowledge, and in minimising the computation time required and making decisions rapidly using AI.
Abstract
Today's manufacturing industry is increasingly turning to simulation, sensing, and data-driven decisions, but these often do not converge into a single workflow. While the Finite Element Method (FEM) can give physically-based predictions of stress, strain, deformation and temperature, models with the high fidelity are slow and expensive to run repeatedly. By comparison, artificial intelligence (AI) absorbs intricacies from computer-simulated and measured data, and upon training, can be programmed to make predictions nearly instantaneously. In this systematic review, following the PRISMA 2020 guidelines, the two approaches are compared in their behaviour when their strengths are combined in the context of predictive risk management in manufacturing. The literature on FEM surrogates and physics-informed learning, AI-enhanced failure mode and effects analysis (FMEA), predictive maintenance, predictive quality and digital twins was synthesised into a 4-category, function-based taxonomy. The evidence demonstrates that the benefits of hybridisation do not lie in replacing simulation with AI, but rather in ensuring predictions are more relevant and reliable using physics-based knowledge, and in minimising the computation time required and making decisions rapidly using AI. Common constraints are: limited data sets, poor cross-domain validation, lack of quantification of uncertainty at the extremes of the training set, poor integration with live production systems; for industrial economies, these are further exacerbated by infrastructure issues, skills issues, and costs. The review thus highlights the gap between the success of lab demonstrations and strong plant-scale deployment. Further research and development is needed in the areas of establishing common benchmarks, risk-aware and explainable hybrid models, multi-fidelity data sets, low-cost deployment routes, and digital-twin models that properly connect simulation, sensing, risk assessment, and operational control.
