Artificial neural networks are widely used in different types of laser technologies. However, traditional accelerators based on the pipeline deployment architecture cannot manage various back propagation (BP) neural networks needed for different laser calculation tasks, such as the extraction of laser welding parameters and laser-induced breakdown spectroscopy analysis. Based on the Xilinx PYNQ-Z2 development platform, a configurable accelerator architecture-based BP neural network for laser welding technologies is designed and implemented herein. By introducing the configurable accelerator architecture and the interconnection of multiplexing operation units, the hardware circuit can be fitted to various BP network structures and the accelerator shows flexible configurability. Furthermore, the data reading method based on a multilevel cache structure is adopted, which addresses the bottleneck of reading speed. Experimental results show that the proposed accelerator can efficiently accelerate the BP neural network with various types of neurons. Compared with the embedded processor platform, the typical network operation performance of the proposed accelerator improves by 10.5 times on average and the large network operation performance with more than 100 neurons improves by 56.4 times on average. The proposed accelerator is superior to the general accelerator, which is realized on the same development platform.