The main design requirement for any high-performance structure is minimal dead weight. Producing lighter structures for aerospace and automotive industry directly leads to fuel efficiency and, hence, cost reduction. For wind energy, lighter wings allow larger rotor blades and, consequently, better performance. Prosthetic implants for missing body parts and athletic equipment such as rackets and sticks should also be lightweight for augmented functionality. Additional demands depending on the application, can very often be improved fatigue strength and damage tolerance, crashworthiness, temperature and corrosion resistance etc. Fiber-reinforced composite materials lie within the intersection of all the above requirements since they offer competing stiffness and ultimate strength levels at much lower weight than metals, and also high optimization and design potential due to their versatility. Braided composites are a special category with continuous fiber bundles interlaced around a preform. The automated braiding manufacturing process allows simultaneous material-structure assembly, and therefore, high-rate production with minimal material waste. The multi-step material processes and the intrinsic heterogeneity are the basic origins of the observed variability during mechanical characterization and operation of composite end-products. Conservative safety factors are applied during the design process accounting for uncertainties, even though stochastic modeling approaches lead to more rational estimations of structural safety and reliability. Such approaches require statistical modeling of the uncertain parameters which is quite expensive to be performed experimentally. A robust virtual uncertainty quantification framework is presented, able to integrate material and geometric uncertainties of different nature and statistically assess the response variability of braided composites in terms of effective properties. Information-passing multiscale algorithms are employed for high-fidelity predictions of stiffness and strength. In order to bypass the numerical cost of the repeated multiscale model evaluations required for the probabilistic approach, smart and efficient solutions should be applied. Surrogate models are, thus, trained to map manifolds at different scales and eventually substitute the finite element models. The use of machine learning is viable for uncertainty quantification, optimization and reliability applications of textile materials, but not straightforward for failure responses with complex response surfaces. Novel techniques based on variable-fidelity data and hybrid surrogate models are also integrated. Uncertain parameters are classified according to their significance to the corresponding response via variance-based global sensitivity analysis procedures. Quantification of the random properties in terms of mean and variance can be achieved by inverse approaches based on Bayesian inference. All stochastic and machine learning methods included in the framework are non-intrusive and data-driven, to ensure direct extensions towards more load cases and different materials. Moreover, experimental validation of the adopted multiscale models is presented and an application of stochastic recreation of random textile yarn distortions based on computed tomography data is demonstrated.