A neural network has been developed to simultaneously use multiple seismic measurements, such as mb(Pn) and coda energy, for estimating yields of underground nuclear explosions. Yield estimation results based on a large number of Nevada Test Site events that were recorded at the broadband seismic stations at Elko, Kanab, Landers, and Mina indicate that different length windows (6, 18, and 200 sec) of seismogram energy are as useful as the mb(Pn) measurement when the seismogram is filtered between 0.5 and 2 Hz. Detailed results of yield estimation trends as a function of signal frequency are also obtained. Yield estimation performance using long lapse time coda energy above 2 Hz is significantly limited by the signal-to-noise ratio (SNR). Performance of the neural network with both mb(Pn) and energy information as input does not significantly exceed the performance of a conventional linear regression analysis that is based only on mb(Pn). Properties of the coda signal as a function of frequency and SNR that might ultimately lead us to the application of a full seismogram with a neural network for accurate seismic yield estimation are discussed.