A comparison between numerical, neural network, and hybrid modelling approaches to simulate spring flow from a karst catchment in northwest Ireland using long-term hydrological data

The study was conducted at the Manorhamilton karst spring, located in County Leitrim, northwest Ireland. This spring represents a typical karst system within the Carboniferous limestone landscape of the region, characterized by highly heterogeneous subsurface flow paths and rapid hydrological responses. This research aims to simulate spring discharge using five years of hydrological data and compare the performance of ten modelling approaches. These include a physically based numerical pipe network model (developed in InfoWorks ICM 2025.1), three neural network (NN) models, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Nonlinear Autoregressive with Exogenous Inputs (NARX), and six hybrid numerical-NN models. Two hybrid strategies were used: Residual Error Correction (REC) and Sequential Combination (SC). All NN models outperformed the numerical model in reproducing karst spring discharge time series. However, the hybrid models, particularly the LSTM+PN+SC variant demonstrated superior accuracy, stability, and generalization across flow conditions. This highlights the added value of combining physical system understanding with deep learning for modelling complex karst systems. The findings offer pros and cons of applying multiple NN models and hybridisation techniques for managing and forecasting groundwater flow in similar Irish karst landscapes. • Modelling of karst spring flow is a challenging task for its complex geological formation. • Numerical, neural, and hybrid models were applied to simulate Irish Karst spring flow. • Hybrid models have been applied with two distinct approaches. • Hybrid numerical-LSTM (SC) gave best results: NSE 0.745/0.68, KGE 0.83/0.73, PBIAS 1.41 %/3.37 % (train/predict).