Engineers are often faced with attempting to characterize data from an experiment without having complete knowledge of the experiment itself or the complete physics of the problem. There may also be limited access to data due to proprietary and/or classification issues. In such cases, developing a physics-based, first-principle model may not be possible. Nevertheless, there may be a need to produce an acceptably precise and/or accurate model that best-approximates the original data. We present an updated procedure for stochastic model development and parameter estimation based upon data-driven principles. Randomly-developed training data is used to demonstrate the newer stochastic methodologies. Improved predictive capability is achieved compared to previously reported results. The model is a suitable surrogate for prediction and enables efficient analysis, integration, and optimization work to proceed without full knowledge of the physics of the experiment. The data-driven stochastic methodology is applicable to all areas of science, engineering, and mathematics.
展开▼
