This paper presents a new rule based systematic screening procedure for synthesis of separation systems. Rules are represented as trapezoidal numbers, enabling partial pattern matching and graded outcome of rules containing continuous variables. Instead of performing conventional conflict resolution on a rule set in order to determine which rule to execute next, it is here proposed simultaneously to evaluate the graded outcome of all rules in the rule base and feed the result of this evaluation to a predictive best-first search mechanism. Clean separation between pattern matching and search is thus achieved. An admissible prediction function has been found, making it possible to guarantee that the search algorithm terminates with optimal solutions when the sharp split search space is restricted to simple separators. The proposed approach is based in part on qualitative knowledge about separation problems. The search algorithm has, therefore, been designed to produce a user specified number of alternative best solutions, which may subsequently be evaluated in more detail using more rigorous models. Examples are included to illustrate the suggested approach.