Proxy and Machine-Learning Models

Too Many Runs

A full simulation can take hours, yet optimization and uncertainty studies need thousands of evaluations. The cost of the simulator is the bottleneck.

Fit a Fast Surrogate

A proxy, or surrogate, breaks it: run the real simulator a handful of times, fit a fast model, a response surface or a neural network, to those samples, and query the cheap proxy instead. Accuracy rises with the number of training runs.

Spend Runs Where It Matters

The art is spending just enough expensive runs to make the proxy trustworthy where the decision is sensitive, often around an optimum. The simulator becomes the teacher; the proxy does the millions of queries, and machine learning is increasingly that proxy.