Process synthesis methods enable the determination of unit operations and their interconnection into a process flowsheet, with associated design and operating parameters, and responding to given objectives. Modern methods are optimization-based, using for example Mixed Integer Non-Linear Programming (MINLP) formulation to optimize a process superstructure. Finding an adequate definition of the search space is a non-trivial problem in such approaches, especially when the number of possible combinations is high due to the process complexity, and is mostly driven by expertise (e.g. heuristics). Consequently, an inductive bias is intrinsically introduced due to restriction of a limited search space, such as the choice of a superstructure representing a limited set of process alternatives. In this work, an evolutionary method is proposed to generate several process architectures based on a set of available unit operations (and associated models) as elementary building blocks. The procedure is here called ab initio process synthesis since it does not require any pre-defined process structure. The developed method relies on the use of an Evolutionary Programming (EP), mimicking natural evolution at species-level, for the automatic construction of a process by using mutation operators to choose, assemble and connect elementary building blocks (i.e. unit operations). A Non-Linear Programming (NLP) is used for process evaluation, by simultaneously solving balances and optimizing process degrees of freedom. The method is implemented in a newly developed tool called PSEvo (Process Synthesis by Evolution). An application to a typical reaction-separation problem is presented, using various problem definitions and evolution control parameters, which demonstrates the method capability to generate optimal processes. The possible uses and the challenges of ab initio process synthesis are finally discussed. (c) 2018 Elsevier Ltd. All rights reserved.