Real-time gamma-neutron discrimination with a trainable polynomial kernel

This paper presents an implementation of gamma-neutron pulse shape discrimination by a support vector machine polynomial decision function in a field-programmable gate array. The training is carried out on a conventional computer using widespread Python libraries. The hardware architecture is designed to allow parameter changes on demand, enabling tuning of hyperparameters and kernel coefficients without requiring re-synthesis. A cubic kernel is compared against a linear kernel which was developed alongside it for non-biased comparison. Both are designed to be viable for real-time classification. The particularities of the designs are explored. The cubic kernel makes use of two stand-alone state machines to keep the sequential data pipelined without interference between the sampled pulses. The results show the trade-off between separation quality, numerical accuracy and physical on-board requirements of the implementations. The separation quality is demonstrated on two datasets, one with a noticeable overlap, to assess any benefits the cubic kernel may bring.