NEURAL NETWORKS APPLIED TO GAAS PROCESS CONTROL

THE OBJECTIVE OF THE PHASE I RESEARCH EFFORT WILL BE TO DEMONSTRATE THE FEASIBILITY OF USING GAAS CRYSTAL RADIUS ESTIMATES, OBTAINED BY APPLYING NEURAL NETWORK TECHNIQUES TO PROCESS VIDEO IMAGERY, FOR REAL TIME CONTROL OF THE LIQUID-ENCAPSULATED CZOCHRALSKI (LEC) PROCESS. A CRYSTAL RADIUS ESTIMATE IS ESSENTIAL FOR IMPROVED PROCESS CONTROL BECAUSE THE CURRENTLY UTILIZED MEASUREMENT HAS AN ANOMALOUS, NON-MINIMUM PHASE BEHAVIOR THAT LIMITS CONTROL SYSTEM PERFORMANCE. NEURAL NETWORK TECHNIQUES ARE APPROPRIATE FOR COMPUTING THIS ESTIMATE BECAUSE OF THE COMPLEXITY OF THE CRYSTAL IMAGES AND THE AVAILABILITY OF TRAINING DATA. FEASIBILITY WILL BE DEMONSTRATED BY DIGITIZING VIDEO IMAGERY AND CORRESPONDING RADIUS MEASUREMENTS OBTAINED FROM THE GE CORPORATE RESEARCH AND DEVELOPMENT CENTER TO PROVIDE TRAINING AND TEST DATA SETS, SELECTING ONE OR MORE CANDIDATE MULTI-LAYER PERCEPTRON (MLP) MODELS, TRAINING THE MODEL(S) TO PERFORM THE NONLINEAR MAPPING BETWEEN THE (PREPROCESSED) IMAGES AND THE CRYSTAL RADIUS, AND EVALUATING THE PERFORMANCE OF THE RADIUS ESTIMATOR FOR DATA NOT INCLUDED IN THE TRAINING SET. ESTIMATION ACCURACY WILL BE ASSESSED BASED ON THE REQUIREMENTS FOR CLOSEDLOOP CONTROL. ASSUMING THAT ESTIMATION ACCURACY IS SATISFACTORY, THE ABILITY TO ACHIEVE REAL-TIME PROCESSING WILL ALSO BE EVALUATED. IMPROVED, INTELLIGENT PROCESS CONTROL OF THE LEC PROCESS CAN REDUCE MATERIAL WASTAGE BY PROVIDING TIGHTER RADIUS CONTROL AND MORE RAPID DETECTION OF THE ONSET OF MULTI-CRYSTALLINE GROWTH. IT CAN REDUCE REQUIREMENTS FOR OPERATOR MONITORING AND POTENTIALLY PROVIDE MATERIAL WITH IMPROVED ELECTRICAL PROPERTIES. ALTHOUGH THE FOCUS OF THE RESEARCH WILL BE ON GROWTH OF GAAS CRYSTALS, THE TECHNOLOGY DEVELOPED WILL BE APPLICABLE TO OTHER MATERIALS AS WELL, E.G., INDIUM PHOSPHIDE.