| In the field of semiconductor manufacturing, the paradigm shift from lambda scaling design rules, to pattern dependent design rules, caused by sub-quarter micron critical dimensions, requires the introduction of sophisticated TCAD and simulation techniques capable of modeling several heterogeneous fabrication processes and optimizing the overall system. The novel optimization technique of Process Metamodeling has been developed and successfully implemented, by constructing neural-network based models of given physical models, representing semiconductor fabrication processes. This second level of abstraction provides both a direct metamodel of a process, which allows the computation of one or several process responses with a speed-up of two orders of magnitude, with respect to standard simulation, and a (totally or partially) inverted process metamodel, which identifies the mapping between process responses and selected process control parameters. This mapping is used in the solution of the parameters extraction (or model calibration) problem. We have demonstrated that, by using a carefully chosen metamodel, it is possible to determine values for unknown (or difficult to estimate) model parameters, which bring a simulated response to overlap, almost identically, with its corresponding experimental one. Several case studies are presented for both direct and inverse metamodels. A complete optimization is also performed for a PEB process for chemically amplified resist. Applications are in the field of Lithography, because of its inherent criticality, with respect to the whole semiconductor manufacturing chain. A novel PEB 3D simulator (which includes photoacid-loss, diffusion and deprotection reactions) was developed as a benchmarking software. Finally, comparisons with other optimization techniques (simulated annealing, genetic algorithms) show that fewer simulation runs are required by process metamodeling for the solution of both optimization and parameters extraction problems. |
