Multi-field numerical model and LSTM-based neural networks for thermal field predictions of concrete-filled steel tubes

Concrete-filled steel tube (CFST) columns are widely applied in long-span bridges due to their high strength, ductility, and construction efficiency. However, in large-diameter CFST members, early-age hydration heat may induce excessive temperature rise and thermal cracking, threatening structural integrity. This study integrates in-situ measurements, multi-field finite element (FE) modeling, and Long Short-Term Memory (LSTM) neural networks to predict thermal behavior of CFST members. Two full-scale columns (2.1 m diameter) and several scaled specimens were tested to record hydration-induced temperature and strain evolution. A chemo-thermal-mechanical coupled FE model was developed and validated against experimental results. Parametric studies revealed the influence of water-to-cement ratio, cement dosage, hydration heat release, and CFST diameter on core temperature evolution. Furthermore, an LSTM network trained on FE-simulated datasets accurately predicted both temperature history and maximum core temperatures, with 99.4% of predictions within 5% relative error. Compared with existing FE–AI hybrid approaches, the novelty of this study lies in the large-diameter CFST range (>2 m), the explicit coupling of chemo-thermal-mechanical fields, and the systematic parameterization of the LSTM training database. The proposed framework provides a reliable and efficient tool for design optimization and risk mitigation in large-scale bridge engineering.