Computer-aided methodology for analyzing IC process and device manufacturability

Manufacturability of integrated circuits depends on the immunity of the design to process fluctuations. Statistical device and circuit simulation can be used to estimate the variation of device and circuit characteristics due to process variations (i.e. the device and circuit manufacturability), and check manufacturability sensitivity to nominal performances in order to guarantee maximally manufacturable and optimal performances. The challenge for efficiently accomplishing the estimation and optimization lies in the difficult task of extracting maximum information about the IC process.;In this dissertation, we present an adaptive divide-and-conquer methodology which can handle these difficulties by capturing the local linearity of the response surface adaptively with an adaptive partitioning scheme. This automated algorithm is based on a piecewise linear statistical modeling approach which can fit the highly nonlinear relations between device/circuit performance functions and a large numbers of process variables, which traditional polynomial regression approaches cannot handle. The resulting models are then used as cheap surrogates for circuit and device simulation in Monte Carlo estimation of the manufacturability as well as in optimization of performances with an acceptable manufacturability and performance variability.;Using the fact that a highly nonlinear performance function can be approximated to arbitrary accuracy, given a sufficiently fine partition of the domain defined by process disturbances and a sufficient number of simulations, the proposed methodology efficiently and recursively partitions the disturbance space into several regions, each of which is then modeled by a linear function. These regions are kept as large as possible provided that they can be modeled accurately by a linear function. Linear models are used because they are less sensitive to the dimension of the disturbance space compared to higher order polynomial models. As the dimension of the problem increases, we only need to increase the number of regions in the partition, rather than the number of terms in the model, which increases exponentially with the order of the model. The distribution of disturbances and the closeness of the regions to the boundary of the feasible region in the disturbance space are used to indicate the importance of regions, since such regions have the greatest impact on estimates of manufacturability. Combined with this criteria for determining the importance of regions, this methodology allows us to refine and resample only the critical regions. Hence, we adaptively refine parts of the model according to their impact on manufacturability and we model different regions with different accuracies. In this way, we reduce the cost of simulation in modeling functions. The optimization of device design, which is achieved by optimizing a composite response (CR) which incorporates three considerations into account simultaneously: the maximal manufacturability, minimum performance variability, and optimal nominal performances, can be transformed into a simple linear programming problem, whereas optimizing the CR using traditional response surface methods, which can be a highly nonlinear and high dimensional optimization problem, requires sophisticated numerical methods.;In this dissertation, the author has also investigated the feasibility and value of using TCAD as an aid in process optimization. (Abstract shortened by UMI.).