Modeling the plasma enhanced chemical vapor deposition process using neural networks and genetic algorithms

Abstract

Silicon dioxide films deposited by plasma-enhanced chemical vupor deposition (PECVD) are useful as interlayer dielectrics f o r metal-insulator structures. In this study, PECVD modeling using neural networks and genetic algorithms has been introduced. The deposition process wiis charucterizea! via a fractional factorial experiment, and data from this experiment were used to train fredTfonvard neural networks using the error back-propagat ion ulgorithm. The networks were optimized to minimize both learning and prediction error: The optimal neural process models were then used for recipe synthesis to generate the proper deposition conditions to obtain specific ,film properties. Th.e response surfaces of the neural process models were explored using genetic ulgorithms, and the performance of this procedure was evaluuted by comparing the deposition conditions indicated by the genetic algorithms w i f h the neurul process model predictions. .0 IntroductionThe PECVD of Si02 in a SiH4/N20 gas mixture yields films with excellent physical properties and useful as interlayer dielectrics for metal-insulator structures such as multichip modules [ 1-21. These properties are determined by the nature and composition of the plasma, which is controlled by the deposition variables involved in the PECVD.However, due to the complex nature of a plasma, it is difficult to quantify the exact causal relationship between controllable deposition conditions and film properties.Although empirical models for plasma-based and other semiconductor processes lhave been developed using statistical response surface moldels 131, process models derived from neural networks have recently been shown to exhibit superior performance in both accuracy and predictive capability [4]. In this paper, accurate process models for the PECVD process are obtained using neural nets, and these models are used to characterize the PECVD of Si02 films deposited under varying conditions. To do so, we have performed a z5-' fractional factorial experiment with three center-point replications [ 5 ] . Data from these 1 Y experiments was used to develop neural process models describing eight output responses: deposition rate, refractive index, permittivity, film stress, wet etch rate, uniformity, silanol concentration, and water concentration.Feed-forward neural networks were then trained on this data using the error back-propagation (BP) algorithm. The development of optimal neural process model is complicated by the fact that back-propagation neural networks contain several adjustable parameters whose optimal values are initially unknown. These include structural parameters (such as the number of hidden layer neurons) as well as BP learning parameters (i.e. -learning rate, momentum, and training tolerance) [61. The effect of these factors on network performance was also investigated, and parameter sets which minimized the training and prediction error of the PECVD models were determined.A recipe synthesis procedure was performed using the optimized neural network model to...

Use of Neural Network to Characterize a Low Pressure Temperature Effect on Refractive Property of Silicon Nitride Film Deposited by PECVD

“…This is mainly attributed to the observation that the temperature affected the RI almost by the same amount irrespective of the N flow rates. The compensation is defined as CRI RI (7) RI RI RI (8) where the subscripts 0 and 1000 represents the flow rates 0 and 1000 sccm, respectively. It should be noted that both and in (8) were measured while setting other process parameters to their medium levels of the experimental ranges shown in Table I.…”

“…The compensation is defined as CRI RI (7) RI RI RI (8) where the subscripts 0 and 1000 represents the flow rates 0 and 1000 sccm, respectively. It should be noted that both and in (8) were measured while setting other process parameters to their medium levels of the experimental ranges shown in Table I. Under these plasma conditions, the RI decreased from 2.19 to 2.12 by about 0.07 as the N flow rate increased from 0 to 1000 sccm.…”

“…Another alternative is to use an empirical approach such as a neural network. Previously, neural networks have Manuscript been applied to a variety of plasma-driven processes [8]- [13]. The use of a neural network is advantageous in that it does not accompany any assumptions, and that the predictive model can quickly be constructed along with fast predictions.…”