dMoT v1.0: A Differentiable Library of 36 Conceptual Hydrological Models for Large-Sample Modelling
Abstract. Conceptual hydrological models represent alternative hypotheses of catchment behaviour, yet their diverse numerical implementations complicate gradient-based, large-sample modelling. We present dMoT v1.0, a PyTorch-based library implementing 36 established lumped rainfall-runoff models within a common differentiable framework. Drawing on MARRMoT formulations, dMoT retains model-specific storages, fluxes, and routing processes while adopting explicit sequential state updates and GPU-parallel computation. We evaluate numerical reliability, automatic differentiation, computational efficiency, and scalability. Water-balance tests met their prescribed tolerances, and automatic-differentiation gradients agreed with finite-difference estimates under the evaluated conditions. Compiled GPU execution achieved a median 2.89-fold forward-execution speed-up over single-threaded CPU execution across the 36 models. We further demonstrate independent calibration and differentiable parameter learning across 531 CAMELS-US catchments. The two approaches achieved median model-level test scores of 0.624 and 0.612, respectively, using a numerically stabilised Kling-Gupta efficiency metric on temporally held-out data from the training catchments. These results establish dMoT as a computational framework for applying contrasting conceptual model structures under different parameter-estimation strategies. By combining a documented model library with automatic differentiation, numerical verification, and scalable execution, dMoT supports reproducible large-sample hydrological modelling and systematic comparisons of model structures and learning approaches.