The thermal modelling of metal additive manufacturing processes such as laser powder-bed fusion (LPBF) involves different time- and length-scales. In the micrometre-scale vicinity of the process zone, high gradient temperature profiles develop which transition at rapid rates (106 K/s) with the moving laser spot. Meanwhile, the centimetre-size component experiences overall temperature increase that evolves at a much slower rate (1 K/min). Application of classical numerical approaches such as the finite element (FE) technique to solve this multiscale problem faces computational efficiency challenges which prevent extension to real size parts. In this work, a machine learning framework has been developed which explores the outcome of a few FE simulations to learn the physical nature of the heat transfer process during LPBF, and later uses the laser scan pattern and overall heat accumulation in the component to predict the detailed temperature profiles at reasonable computational costs. The accuracy and computational efficiency of the proposed approach are benchmarked based on equivalent FE models.