TimoDS: A comprehensive and synthetically generated dataset of Timoshenko steel beams with general elastic supports and responses to elementary unit load cases

<title>Abstract</title> <p>This paper presents TimoDS, a synthetically generated dataset of 60000 Timoshenko steel beam instances with general elastic supports, together with a baseline machine learning experiment demonstrating its utility for data-driven structural mechanics. Each instance corresponds to a fully defined linear static beam problem, characterised by a beam length, a European steel cross-section from the IPE, HEA or HEB profile families, and twelve elastic support stiffness parameters spanning the full continuum from free to quasi-clamped boundary conditions. Six elementary unit load cases, comprising three unit point loads and three unit point moments, are provided independently for each instance, enabling the reconstruction of responses to arbitrary loading distributions via superposition. For each problem, the spatial distributions of kinematic quantities, internal forces and strain energy are sampled at 21 equally spaced positions along the beam axis, yielding an output vector of 253 quantities per instance. The dataset was generated through a fully reproducible pipeline combining Latin Hypercube Sampling with a fixed random seed and an automated parametric graph implemented in Grasshopper and Karamba3D. A feedforward neural network surrogate trained on a single load case demonstrates accurate reconstruction of response fields and confirms that the superposition principle is satisfied by the trained model on out-of-sample loading scenarios, establishing the dataset as a viable benchmark for surrogate modelling, physics-informed learning and operator learning in structural engineering.</p>