Using Neural Networks to Accelerate TALYS-2.0 Nuclear Reaction Simulations

TL;DR

It is shown in this work that an artificial neural network can serve as a surrogate model to successfully predict TALYS-2.0 outputs within this domain of input parameters.

Abstract

Recent efforts to improve the predictability of TALYS-2.0 calculated charged-particle residual product cross sections have focused on adjusting parameters related to the optical model potential and pre-equilibrium process. Although adjusted TALYS-2.0 outputs show marked improvements in agreement with experimental data over the default parameters, the procedure is generally time-consuming due to the need for sequential TALYS-2.0 calculations. Since the models and model parameters must be defined and constrained prior to adjustment, we show in this work that an artificial neural network can serve as a surrogate model to successfully predict TALYS-2.0 outputs within this domain of input parameters. No practical differences were observed in the trained model's performance between uniform random, Latin hypercube and Sobol sequence sampling for generating the training datasets. Once validated, trained neural network models were used to adjust TALYS-2.0 nuclear model parameters, where a multi-parameter fitting procedure was not only feasible but optimal for this process. The neural network approach is >1000x faster at generating residual product cross sections than using TALYS-2.0 directly, and a high-fidelity surrogate model could be implemented with about 1500 TALYS-2.0 files to achieve adjusted cross sections comparable to the previous publication.

Figures (13)

• Figure 1: Comparison between the NN surrogate model (red solid) and standard cubic interpolator (red dotted) after training with 6x6x6 TALYS-2.0 files. The cubic interpolator incurred substantially larger errors than NN for generating the test data (red circle). For reference, the reaction channel shown is natCu(p,x)51Cr, and three separate parameter sets used for training both models are plotted as black hollow markers for illustration.
• Figure 1: Comparison of the previous approach by Morrell et al. morrell_measurement_2024 and the proposed approach in this work for adjusting TALYS-2.0 nuclear model parameters.
• Figure 2: Different metrics used to evaluate the performance of the surrogate model trained using files sampled from uniform (blue triangle), Latin hypercube (LHC, orange circle) or Sobol sequence (green square) sampling. All three sampling methods are generally comparable and show a general trend of improved performance with increasing training dataset size.
• Figure 2: Actual (TALYS-2.0) vs predicted (NN model) plots for each of the sampling methods and respective number of training files used to train the NN model for 139La(p,x). The TALYS-2.0 cross section data were taken from 25k uniform random sampled files not used for training. Due to the large amount of data, only 750k random cross sections were selected for these plots.
• Figure 3: Resulting excitation functions after the N-D parameter adjustment procedure using the NN model trained with either 2048 files from Sobol sequence (orange dashed) or 1536 files from uniform random sampling (pink dashed). The respective TALYS-2.0 cross sections (dash dotted), experimental data (black asterisk), default TALYS-2.0 (black solid) and previously adjusted output by morrell_measurement_2024 (red dotted) are shown for comparison.