USING MEASURED DAMAGE PARAMETERS TO PREDICT THE RESIDUAL STRENGTH OF IMPACTED COMPOSITES - A NEURAL-NETWORK APPROACH

The poor performance of composite materials under transverse quasi-sta tic and impact loading is of major concern in their application as pri mary load carrying components in advanced structural applications. The se materials sustain substantial internal damage in the form of matrix cracking, delaminations and fiber fracture. The present paper reports the results of a modeling exercise which used neural networks as a to ol to predict the loss in residual strength resulting from localized d amage in impacted laminates. Several measured fiber fracture parameter s, as well as matrix damage areas, obtained from damaged laminates wer e used as inputs. Neural networks were used to identify those damage p arameters that were essential for effective residual strength predicti on. Development of the neural network models was performed using exper imental data from specimens fabricated from the Fiberite IM7/977-2 mat erial system which were first damaged via quasi-static contact loading , and then loaded in tension to failure. Data obtained from specimens tested for residual strength after impact were used to test the model' s generalization capability. A pruning study was also conducted to det ermine an optimally connected, robust neural network model that genera lized better than conventional fully connected feedforward networks. T he predicted strength values were found to be in very good agreement w ith those obtained from experiments indicating the suitability of neur al networks in this application.