Machine learning-based GA optimization of lumbar intervertebral disc FE models: Comparison with RSM and experimental data.

Finite element (FE) models are widely used to investigate the mechanical behaviour of the human Intervertebral Disc (IVD), but their predictive capability strongly depends on the accurate calibration of multiple material parameters. This study proposes a surrogate-assisted optimization framework that integrates finite element modelling, machine-learning (ML) regression models and Genetic Algorithms (GA) to identify the material parameters governing a healthy lumbar intervertebral disc. A parameterized L3-L4 intervertebral disc FE model was constructed using eleven material parameters, which were systematically varied through a two-level fractional-factorial Design of Experiments (DoE). This yielded 128 finite element simulations under six standard loading scenarios, producing six stiffness responses and nine bulge deformations that were used to train surrogate regression models for all mechanical responses. The surrogate models showed excellent predictive capability for most responses, enabling efficient exploration of the high-dimensional parameter space without additional FE simulations, although slightly lower accuracy was observed in some stiffness-related outputs due to their higher complexity. GA-based optimization produced parameter sets that more accurately reproduced experimental reference data than those obtained using a previous Response Surface Methodology (RSM) approach. This improvement was reflected in a 15-20% reduction in normalized mean absolute error, with GA achieving values of 0.229-0.238 compared with 0.278-0.280 for RSM. The largest improvements were observed in bulge-related responses, although stiffness predictions also benefited from the GA approach. Overall, the results demonstrate that combining FE modelling with surrogate machine learning models and genetic algorithms provides a powerful and computationally efficient strategy for calibrating lumbar IVD FE models and improving agreement with experimentally observed mechanical behaviour. The proposed framework enhances accuracy, reduces computational cost and facilitates the development of more realistic biomechanical simulations with potential applications in spinal implant design, pre-clinical testing and patient specific modelling.