Purpose: To train a neural network to accurately predict alpha particle microdosimetric parameters, namely, the average and average square single-hit specific energy,
Methods: A two-layer feed forward neural network consisting of 4 inputs and 2 outputs was adapted from MATLAB and trained on sufficiently large previously published microdosimetric data. The data spans the common range of alpha-particle energies (3.97 – 8.78 MeV), nuclear radii (2 – 10 microns), cell radii (3 – 18 microns), and several source-target configurations. The mean square error (MSE) was used to measure the performance of the function fitting neural network. The random data partition for percent training/validation/testing and the number of nodes in the hidden layer were chosen to minimize MSE while ensuring the network could generalize well. The output layer transfer function was chosen to restrict outputs to physically meaningful positive values. The Levenberg-Marquardt training algorithm was implemented. A neural network of 8 hidden nodes and a 70%/15%/15% data partition was trained on 672 data points. The network can be exported and shared for tests on independent data sets and calculations of
Results: The network predictions and target values showed good agreement. The best training performance of MSE = 8.323 x 10⁻⁵ was observed at epoch 153, and training was stopped after satisfying 6 validation checks. The network performed with a final test MSE = 1.306 x 10⁻⁴. The overall target to output correlation was R = 0.99981. The mean percent error was -1.33% for
Conclusion: This trained neural network can produce microdosimetric parameters used for the study of alpha-particle emitters. Further directions include expanding to more realistic configurations and testing the network on independent data sets.