Architecture generation and evaluation are critical points in complex system design. Uncertainties concerning component characteristics and their impact onto overall system performance are often not taken into account in early design stages. In this paper, we propose a Bayesian network (BN) approach for system architecture generation and evaluation. A method relying on Bayesian network templates is proposed in order to represent an architecture design problem integrating uncertainties concerning component characteristics and component compatibility. These templates aim at modeling designers' knowledge concerning system architecture. We also propose an algorithm for architecture generation and evaluation related to the Bayesian network model with the objective of generating all possible architectures and filtering them in view to a defined confidence threshold. Within this algorithm, expert estimations on component compatibilities are used to estimate overall architecture uncertainty as a confidence level. The proposed approach is tested and illustrated on a case study of bicycle design. This first case shows how uncertainties concerning component compatibilities and components characteristics impact bicycle architecture generation. The method is, additionally, tested and implemented in the case of a radar antenna cooling system design in industry. Results highlight the relevance of the proposed approach in view to the generated solutions as well as other benefits such as reduced time for architecture generation, and a better overall understanding of the design problem. However, some limitations have been observed and call for enhancements like integration of designer's preferences and identification of possible trade-offs within the architecture. This method enables generation and evaluation of complex system architecture taking into account initial system requirements and designer's knowledge. Its usability and added-value have been verified on a large-scale system implemented in industry.