Optimal experimental design with fast neural network surrogate models

Designing optimal experiments minimizes the uncertainty of results and maximizes efficient use resources. Herein, machine learning surrogate models approximate coordinate exchange (ACE) algorithm are used to determine experimental designs (OEDs) over large or arbitrarily restrictive design spaces. OED is particularly salient in materials science, where expensive material properties must often be inferred indirectly. The proposed framework demonstrated by finding with which hidden constituent composite can most efficiently from observable outcomes. given an information-theoretic criterion that conditional mutual information between expected To perform tractable optimization, a neural network trained as model mimic physics-based simulation, calculate outcome based on candidate sampled properties. ACE optimize spaces many tests controlled parameters exhaustive search would intractable even model. Using this approach, OEDs consistent those produced heuristic knowledge established best practices found; then larger unavailable examined. All code, data, reproduce work paper available at https://github.com/nasa/OED-with-NN-surrogates.