Machine-Learning Design Optimization and Tomographic Inversion of 2D Fiber Optic Bolometer Array

New inversion schemes are demonstrated to assess the feasibility of a novel 2D fiber-optic bolometer (FOB) array. A recent test showed that the FOB is comparable to a resistive bolometer performance and avoids electromagnetic interference by using Fabrey-Pérot resonator system to encode small temperature changes related to the incoming power. With these advantages and its compact size, an array was designed using a Bayesian global optimization algorithm with CHERAB library. The algorithm efficiently optimized six design parameters with non-linear results. Six synthetic radiation profiles were used to calculate the cost for the optimization. Tomographic inversion methods for the optimized design were developed using an iterative method with regularization and neural networks (NN). The regularizations were higher costs as further away from the separatrix and lower costs along the similar flux values. 20,000 partially randomized synthetic radiation profiles were used to train the NN methods. The sensitivity matrix of the design was used with a measurement as an input for the “only” NN method. A hybrid method used the result of the iterative method with a simple regularization as an input for a NN.

*Work supported by the U.S. DOE (Grant Nos. DE-AC05-00OR22725, DE-FC02-04ER54698).