Non-equilibrium ionization plays a critical role in Z-pinch gas discharge produced plasma (GDPP) EUV source. However, the physics of the processes, plasma and surface discharges produced, magneto-hydrodynamic, photon radiation transport, and plasma-electrode interactions, which lead to EUV emission, is intrinsically complex. Many simplifying assumption are inevitable with numerical simulations, resulting in low-credibility outcomes. With the learning and generalization abilities, artificial neural networks (ANN) have been applied to model and optimize a Z-pinch plasma source, which is characterized with a experimental design at varied operational parameters including electric power input, applied voltage/current, pulse repetition, MPC parameters, electrode geometry, xenon flow rate as well as convention efficiency, EUV source size, radiation power etc.

z h a~g~e e c s .~m a m o t o -~. a c~pIn our studies, EUV radiation from xenon-filled fast capillary Z-pinch discharge has been made. The urgent need for modeling and optimizing such xenon plasma sources requires theoretical efforts in atomic and plasma physics. However, the physics of the plasma-produced processes, which lead to E W emission, are intrinsically complicated. Many simplifying assumption is inevitable with numerical simulation, resulting in suspicious outcomes. ANN can learn complex plasma data both adaptively and nonlinearly without any formation on the causal relationship between the input and output patterns. With the learning and generalization abilities, ANN is utilized to model and optimize Z-pinch piasma source, which is Characterized with a experimental design at vaned operationaJ parameters including electric power input, applied voltageicurrent, pulse repetition, MPC parameters, electrode geometry, Xe flow rate as well as convention efficiency, EUV source size, radiation power etc.A schematic of the 2-Pinch EUV sources is shown in Figure 1. The EUV radiation from the Z-Pinch plasma was characterized with measurement for the temporal behavior of E W intensity and the pinhole images. It is worthy to be noticed the maximum EUV radiation is the most sensitive to the Xe flow rate and the discharge current. ANN is wed to construct evaluation model for EUV plasma process, the back-propagation (BP) neural network, shown in Figure 2., is employed due to its h g h prediction capability in modeling complex EUV plasma data. The ANN input patterns are represented by the four adjustable process factors, i.e., gas flow rate, charge voltage, discharge current and input energy; the output layer is set to unity neuron, representing EUV power [W/5W2%BW/2pisr]; the hidden layer with five neurons do not interact with the outside world, but assist in performing nonlinear feature extraction on the data provided by the input and output layers. The ANN is trained with BP algorithm, a total of 12 experiments were conducted to produce training patterns, and additional five experimental pattems were used as test data for EW evaluation. Figure 3. Shows the EUV evaluation by use of ANN. For comparison, the actual measurement values are also shown in the Figure 3. As seen, the evaluation vahes of EUV power via ANN are highly consistent with the corresponding actual measurements.NevertheIess, there exist some differences between experimental and actual measurements, so that the model is to be fiuther optimized such as ANN structure including the numbers of ANN layers, of each hidden layers as well as the improvement of training algorithm. In addition, the training pattems are to be collected again. For comparison, the dependences of EUV power on the gas flow rates and discharge currents, respectively, are shown in Figures 4 and 5. Whereat, the generalization capabilities of ANN are exploited to intelpolate between training data, thus to overcome system parametric uncertainties. Therefore, the behaviors of EUV power as a...