Researchers from the University of Pittsburgh have developed a scalable and adaptive sampling method for surrogate modeling.
Description
Surrogate modeling to approximate results of experiments and analytical models are commonly used in physics and engineering to speed up and simplify engineering simulations. Adaptive modeling, whereby information from one iteration is used to generate sample data for the next iteration is not suitable for complex models due to exponential increase in the required sample size. This novel approach uses active machine learning where artificial neural network models iteratively apply random sampling for the generation of testing samples and training surrogate models. The model is scalable for high-dimensional inference spaces without the need for high computational power.
Applications
• Computational modeling
• Engineering modeling
• Machine learning
Advantages
Traditional and adaptive sampling techniques for machine learning models often suffer from the “curse of dimensionality” whereby the requirement for sampling points increases exponentially as the complexity of the problem being solved increases. Such limitations often make the modelling of complex problems unfeasible.
This new approach is designed to overcome these challenges through testing the model at each iteration on a set of random points. Based on the results and testing points, new training points generated as a subset of a pre-defined, full-factorial regular mesh are added, with constant refinement of the full-factorial sample at each iteration. This approach allows the scalability of sampling with the dimensionality of the inference space than other adaptive sampling techniques, ensuring the development of high dimensional neural network models without the need for very large computational power.
Invention Readiness
An algorithm has been developed and evaluated on several illustrative examples with known analytical solutions and a practical case study with a neural network surrogate model. For a 4D inference space, this novel sampling method had an order of magnitude lower error and a sample size 20 times smaller than that required by traditional one-shot sampling in a 6D inference space.
IP Status
https://patents.google.com/patent/WO2024151535A1
