<p>Flood susceptibility mapping plays a critical role in flood risk management and spatial planning, particularly in regions frequently affected by hydrological extremes. This study evaluates the performance of hybrid machine learning models that integrate a multi-layer perceptron (MLP) neural network with three optimization algorithms: Artificial Bee Colony (ABC), Elephant Herd Optimization (EHO), and Differential Evolution (DE). The models were applied to the Putna River basin, Romania, using a flood inventory of historical events and fourteen flood-conditioning factors derived from topographic, hydrological, geological, and land-use data within a GIS environment. Model performance was assessed using accuracy, precision, recall, F1-score, and the area under the ROC curve (AUC). Among the tested approaches, the DE–MLP model showed the most robust and generalizable performance, achieving an accuracy of 0.982 and an AUC value of 0.985 on the testing dataset. The ABC–MLP and EHO–MLP models also produced reliable results but with slightly lower predictive capability, while the single MLP model exhibited comparatively reduced performance. Variable importance analysis indicates that slope, elevation, distance to river, and rainfall exert the strongest influence on flood susceptibility in the study area. The results demonstrate that optimization-based hybrid models, particularly DE–MLP, can significantly enhance flood susceptibility prediction and provide valuable decision-support tools for flood risk mitigation and territorial planning.</p>