Abstract: Reliable prediction of the martensite start temperature (M_s) is essential for understanding phase transformations and optimizing heat treatment processes in steels. However, empirical relations commonly used for M_s estimation are often limited by simplified formulations and restricted compositional ranges. In this study, 20 widely used empirical equations are systematically evaluated and compared with an artificial neural network (ANN) model for low-carbon steels (C content < 0.30wt%). A dataset comprising 285 steel compositions with 15 alloying elements was compiled, with 171 samples used for training and 114 for testing. The optimized ANN architecture (15–10–10–1), representing 15 input neurons, two hidden layers with 10 neurons each, and one output neuron, achieved a test-set mean absolute error (MAE) of 65.49 K, representing a 35%–50% reduction relative to the best-performing empirical equations. Sensitivity analysis of the trained ANN model further revealed metallurgically consistent effects of alloying elements on M_s. The results identify the relative contributions of alloying elements and their combined influence on martensitic transformation behavior in low-carbon steels. Experimental validation through quenching-and-partitioning heat treatment produced the expected martensite and retained austenite microstructures.