Proposal of new Combustion Analysis Concepts Based on Learning and Inference Functions.

This paper presents concepts of new combustion analysis methods, and a learning and inference algorithm for the concepts. The analysis methods are based on the following three concepts : (1)constructing a database on the relationships between experimental conditions and combustion characteristics, and predicting combustion performance for a given condition ; (2) extracting the physical relationships among the combustion factors ; (3) enabling deficiencies in measurements and numerical simulations to compensate each other by adjusting uncertain parameters in the simulation to fit the calculated results to the measurement based on learning and inference functions. In the present study, we attempt to develop a new method of inference which enables analysis of nonlinear multi-variable relationships using a limited number of data. A feasibility study of this method showed that good learning and inference performance were achieved using these concepts.