Advanced neural network model of plasma-driven integrated circuit process data

Plasma processes are key means to deposit or etch thin films during the manufacture of integrated circuits. An advanced model of plasma process data was constructed by applying genetic algorithm to the set of typical training factors combined with multi-parameterized gradients of neuron activation functions. The presented technique was evaluated with plasma etch data, collected during silica etching in CHF3-CF4 inductively coupled plasma. The etch responses measured include Al etch rate, Al selectivity over silica etch rate, silica profile angle, and DC bias. Constructed models demonstrated better predictions over conventional models for all etch responses. The improvement measured by the relative percentage error was even more than 30%. This was also demonstrated over previous models. In consequence, the presented modeling technique can be effectively applied to construct accurate prediction models with a limited set of experimental data.