In competitive industries, a reliable yield forecasting is a prime factor to accurately determine the production costs and therefore ensure profitability. Indeed, quantifying the risks long before the effective manufacturing process enables fact-based decision-making. From the development stage, improvement efforts can be early identified and prioritized. In order to measure the impact of industrial process fluctuations on the product performances, the construction of a failure risk probability estimator is developed in this thesis. The complex relationship between the process technology and the product design (non linearities, multi-modal features...) is handled via random process regression. A random field encodes, for each product configuration, the available information regarding the risk of noncompliance. After a presentation of the Gaussian model approach, we describe a Bayesian reasoning avoiding a priori choices of location and scale parameters. The Gaussian mixture prior, conditioned by measured (or calculated) data, yields a posterior characterized by a multivariate Student distribution. The probabilistic nature of the model is then operated to derive a failure risk probability, defined as a random variable. To do this, our approach is to consider as random all unknown, inaccessible or fluctuating data. In order to propagate uncertainties, a fuzzy set approach provides an appropriate framework for the implementation of a Bayesian model mimicking expert elicitation. The underlying leitmotiv is to insert minimal a priori information in the failure risk model. Our reasoning has been implemented in a software called GoNoGo. The relevancy of this concept is illustrated with theoretical examples and on real-data example coming from the company STMicroelectronics.
