The reliable monitoring of liquid front progression during infusion processes remains a key challenge for the industrial deployment of out-of-autoclave lightweight composite manufacturing. This study investigates the use of ultrasonic guided waves for fluid front localisation in a bespoke liquid-only experimental setup, evaluating data-driven waveform modelling approaches and strategies for bridging the gap between simulated and experimental signals. A refined finite element simulation workflow is developed to reproduce the fundamental modal behaviour of experimental waveforms at substantially reduced computational cost when compared to other previously implemented models. A one-dimensional convolutional neural network trained directly on time-domain signals achieves a mean absolute error of 7.71 ± 2.11 mm across five independent infusion runs, outperforming the baseline energy-based functional approximation. To address the limited availability of experimental data, a conditional Generative Adversarial Network (GAN) is developed to bridge the gap between the simulated and experimental domains. When GAN-adapted synthetic data is combined with a single experimental dataset for training, the prediction error is reduced by 46% compared to training on experimental data alone, from 16.69 ± 0.86 mm to 9.07 ± 1.11 mm. These results demonstrate that waveform-based deep learning models, supported by GAN-adapted synthetic data, offer a promising and data-efficient route for accurate monitoring of resin flow in liquid composite moulding processes.