The propagation of ultrashort pulses in optical fibre plays a central role in the development of light sources and photonic technologies, with applications from fundamental studies of light–matter interactions to high-resolution imaging and remote sensing. However, short pulse dynamics are highly nonlinear, and optimizing pulse propagation for application purposes requires extensive and computationally demanding numerical simulations. This creates a severe bottleneck in designing and optimizing experiments in real time. Here, we present a solution to this problem using a recurrent neural network to model and predict complex nonlinear propagation in optical fibre, solely from the input pulse intensity profile. We highlight particular examples in pulse compression and ultra-broadband supercontinuum generation, and compare neural network predictions with experimental data. We also show how the approach can be generalized to model other propagation scenarios for a wider range of input conditions and fibre systems, including multimode propagation. These results open up novel perspectives in the modelling of nonlinear systems, for the development of future photonic technologies and more generally in physics for studies in Bose–Einstein condensates, plasma physics and hydrodynamics.

超短脉冲在光纤中的传播在光源和光子技术的发展中发挥着核心作用，其应用从光物质相互作用的基础研究到高分辨率成像和遥感。然而，短脉冲动力学是高度非线性的，为了应用目的而优化脉冲传播需要广泛且计算要求高的数值模拟。这给实时设计和优化实验造成了严重的瓶颈。在这里，我们提出了一个解决这个问题的方法，使用循环神经网络来模拟和预测光纤中复杂的非线性传播，仅从输入脉冲强度分布。我们重点介绍脉冲压缩和超宽带超连续谱生成中的特定示例，并将神经网络预测与实验数据进行比较。我们还展示了如何将该方法推广到为更广泛的输入条件和光纤系统（包括多模传播）模拟其他传播场景。这些结果为非线性系统建模、未来光子技术的发展以及更普遍的物理学研究开辟了新的视角，用于玻色-爱因斯坦凝聚、等离子体物理学和流体动力学的研究。