Machine learning method based on cluster dynamics for accelerating the prediction of dislocation loop evolution in irradiated austenitic steel

• A general framework for building surrogate models is developed to accelerate cluster dynamics simulations. • SHAP-integrated visualization enables intuitive parameter ranking and sensitivity analysis for cluster dynamics. • Precise multi-parameter coupled sensitivity analysis is achieved without additional cluster dynamics simulations. • Automated parameter calibration using experimental data and optimization algorithms minimizes simulation-experiment deviations. Cluster Dynamics (CD) is a robust method for predicting irradiation-induced defects, yet its application to high-dose scenarios is hindered by prohibitive computational costs and the challenges of manual parameter calibration. In this work, we propose a computationally efficient surrogate model by applying Bidirectional Long Short-Term Memory (BiLSTM) networks to CD for the first time. The proposed surrogate model accurately predicts the nonlinear evolution of dislocation loops in austenitic steel up to 50 dpa within seconds, bypassing the on-the-fly solution of the governing equations within the CD model. To enhance interpretability of the surrogate model, Shapley Additive Explanations (SHAP) are employed to quantify input feature contributions. The SHAP-based interpretability analysis aligns with physical parameter sensitivity, establishing a novel, computationally efficient method for characterizing the effects of physical parameters on CD outcomes. The model’s superior in-domain prediction performance also enables a refined multi-parameter sensitivity analysis without additional computational overhead. Furthermore, coupling the surrogate model with the NSGA-II multi-objective optimization algorithm enables automatic parameter calibration, effectively minimizing deviations between predictions and experimental statistics. This machine learning-assisted framework achieves seconds-level predictions for high-dose irradiation damage. Coupled with automated optimization algorithms, it establishes a novel paradigm integrating rapid prediction, physics-based interpretation, and automated parameter calibration.