The natural variability of wood properties and its kottiness remains a significant challenge for predicting the load-bearing capacity of timber structures. This study uses the TreeTrace openaccess database to predict investigate the resulting bending strength of 346 Douglas-fir beams by coupling numerical damage modelling with global sensitivity analysis. For each beam, a 2D digital twin is generated from the measurement performed on faces, incorporating local grain angles, density. Knots morphology is detected via the YOLO algorithm. The mechanical behavior is simulated using a non-linear damage model. To limit the high number of finite element simulations, a Gaussian Process Regressormetamodel (Scikit-learn, Python) was trained for each twin, enabling the exploration of over 20,000 parameter combinations. To define the physical boundaries of the simulation, the following table summarizes the ranges of the mechanical variables investigated, which constitute the input domain for the sensitivity analysis. In this initial phase, the study focuses on a subset of 15 beams exhibiting tensile failure. A sensitivity analysis based on Sobol indices was performed to rank the influence of the 6 mechanical parameters listed ''reference du tableau" on the global structural response. Numerical stiffness is precalibrated by updating the experimental Young's modulus . On this subset, preliminary results indicate that under bending loads—specifically for beams with tensile crack failure—the most significant parameters are, on average, the Radial/Tangential tensile strength followed by the Longitudinal tensile strength. Conversely, fracture energies and shear strength show a negligible influence on the overall load-bearing capacity in this specific configuration.