Fast and accurate representation of heat transfer in laser powder-bed fusion of metals (PBF-LB/M) is essential for thermo-mechanical analyses. As an example, it benefits the detection of thermal hotspots at the design stage. While traditional physics-based numerical approaches such as the finite element (FE) method are applicable to a wide variety of problems, they are computationally too expensive for PBF-LB/M due to the space- and time-discretization requirements. Alternatives are sought to lower the computational burden of modelling this process and make part-scale simulations feasible, with machine learning (ML) techniques leading these efforts due to their exceptional flexibility and efficiency. Recently, there has been a growing trend towards integrating physical insights of the studied phenomena in ML workflows to improve their effectiveness. For the presented work, we hypothesized that the moving laser heat source could be treated as a sequence of pulses such that the solution to various scan patterns could be determined based on the thermal response to a laser pulse. First, a base function represented by a feed-forward neural network (FFNN) was proposed to establish the solution for laser scanning over a wide solid block. Next, inspired by the perturbation theory, a second FFNN was introduced to consider the impact of geometrical features on the temperature profiles as a correction to the base solution. The feasibility of training the pair of FFNNs within the proposed 'pulse approach' framework based on a few inexpensive FE simulations, and generalization to larger simulation domains are demonstrated. For a scan pattern not encountered during training, the paired networks are capable of accurately replicating the temperature profiles or history predictions of FE simulations in under one second, showcasing a considerable acceleration by orders of magnitude. The models and scripts used in this study are openly available in https://github.com/HighTempIntegrity/PIAM_Pulse2024.