Integrated ANSYS Transient Thermal Simulation and Machine-Learning-Based Defect Prediction for an L-Shaped Structural Steel Sand Casting

Abstract

This paper presents an integrated framework that couples ANSYS Workbench transient thermal finite element simulation with a Python-based artificial intelligence (AI) pipeline for predicting thermal defects in sand-cast structural steel components. A 558.8 × 558.8 × 254.0 mm two-body assembly (495.45 kg total mass), comprising an L-shaped structural steel casting within a sand mould, was meshed with 171,227 nodes and 37,947 elements and subjected to a step pouring-temperature boundary condition of 1579.4°C with radiative cooling (ε = 0.3, Tₐₘₙ = 28°C) over a 5-second transient analysis comprising 14 substeps. The raw ANSYS outputs — global minimum/maximum temperature and probe heat flux — were transformed via Python feature engineering into 11 physics-based indicators, including a Niyama-criterion proxy, and used to train and benchmark three machine-learning classifiers (Random Forest, Gradient Boosting, SVM-RBF) and a convolutional neural network on a five-class defect taxonomy (Hot_Spot, Porosity, Shrinkage, Non_Uniform_Cooling, No_Defect). The Random Forest classifier achieved 100% test accuracy (93.7% cross-validation accuracy, AUC = 1.000), correctly identifying a persistent Hot_Spot condition at the L-junction throughout the entire analysis window (T_max never falling below 1592°C, 92°C above the steel liquidus). A composite 0–100 risk score, derived from four weighted thermal indicators, declined from 70.2 at t = 0.05 s to 60.3 at t = 5.0 s, tracking the 98.2% decay of probe heat flux. The study demonstrates that physics-informed feature engineering enables highly accurate, interpretable, real-time defect classification directly from coarse ANSYS scalar outputs, and provides quantitative engineering recommendations — chill placement, riser design, reduced pouring superheat, and digital-twin integration — for mitigating the identified hot-spot and thermal-shock risks.

Keywords