Using Neural Networks to Accelerate TALYS-2.0 Nuclear Reaction Simulations

This paper demonstrates that an artificial neural network can serve as a high-fidelity surrogate model for TALYS-2.0, accelerating the generation of charged-particle residual product cross sections by over 1000 times while enabling efficient multi-parameter adjustments to improve agreement with experimental data.

The Gist

Imagine you are trying to bake the perfect cake, but you don't have a recipe. You only have a very complex, slow-moving robot chef (let's call it TALYS-2.0) that can bake a cake, but it takes hours to figure out exactly how much sugar, flour, and heat to use to get the flavor just right.

In the world of nuclear physics, scientists need to predict how atoms react when hit by particles (like protons) to create useful isotopes for medicine and research. The "robot chef" (TALYS-2.0) is the standard tool for this, but it's incredibly slow. If you want to tweak the "ingredients" (nuclear model parameters) to match real-world experiments, you have to ask the robot to bake a cake, check the taste, change an ingredient, and ask it to bake again. Doing this hundreds of times in a row takes days or weeks.

The Problem:
The scientists in this paper wanted to speed up this process. They needed a way to find the perfect "recipe" without waiting for the slow robot chef to bake every single trial cake.

The Solution: The "Smart Apprentice"
The team built a Neural Network, which is essentially a "Smart Apprentice" trained by the robot chef.

1. The Training Phase: Instead of asking the robot chef to bake millions of cakes one by one, they asked it to bake about 1,500 different cakes with random ingredient combinations. They fed all the results (how the cakes turned out) into the Smart Apprentice.
2. The Learning: The apprentice studied these 1,500 cakes. It learned the patterns: "Oh, if I add a little more of ingredient X and less of Y, the cake gets sweeter." It didn't just memorize the cakes; it learned the logic of the kitchen.
3. The Magic: Once trained, the apprentice could predict the outcome of any new ingredient combination in a fraction of a second. It became a "surrogate model"—a fast, fake version of the real robot chef that was 99% accurate but 1,000 times faster.

The Experiment:
The scientists tested this apprentice in three ways:

• The Simple Test: They gave it a recipe with just 3 ingredients. It learned quickly and predicted the results almost perfectly, even better than a standard math calculator they tried to use as a backup.
• The Hard Test: They increased the ingredients to 6, then 17. The math calculator crashed and gave up because the math was too hard. The Smart Apprentice? It didn't even break a sweat. It handled the complexity easily.
• The Real World Test: They used the apprentice to adjust the parameters for a real nuclear reaction (Lanthanum-139). They asked the apprentice to find the perfect settings to match real experimental data.

The Results:

• Speed: The process that used to take days now took minutes. The apprentice was over 1,000 times faster than the original robot chef.
• Accuracy: The "perfect recipes" found by the apprentice matched the real-world data just as well as the slow, manual method, and sometimes even better.
• Flexibility: Because the apprentice is so fast, scientists can now try thousands of different "what-if" scenarios instantly. They can change the goal (e.g., "I want a cake that is less sweet but fluffier") and get an answer immediately, without waiting for the slow robot.

The Bottom Line:
This paper is about teaching a computer to "dream" the results of complex nuclear physics simulations. By training a neural network on a small sample of data, they created a super-fast shortcut. This allows scientists to design better medical isotopes and understand nuclear reactions much faster, turning a process that used to be a slow, tedious slog into a quick, efficient sprint.

In short: They built a "crystal ball" that predicts nuclear reactions instantly, saving them years of waiting time.

Technical Summary

1. Problem Statement

The production of critical isotopes for medical, security, and research applications relies heavily on accurate nuclear cross-section data. When experimental data is unavailable, codes like TALYS-2.0 are used to predict these values. However, optimizing TALYS-2.0 parameters to match experimental data is a significant bottleneck:

• Sequential Processing: The traditional parameter adjustment workflow requires sequential TALYS-2.0 calculations, making the process time-consuming and preventing parallelization.
• Computational Cost: Iterative fitting procedures involving complex nuclear models (e.g., optical model potentials, pre-equilibrium processes) are computationally expensive, limiting the ability to perform high-fidelity sensitivity analyses or global optimizations.
• Need for Surrogates: There is a need for a "surrogate model" that can approximate TALYS-2.0 outputs with high fidelity but at a fraction of the computational cost, enabling rapid parameter tuning.

2. Methodology

The authors propose using Artificial Neural Networks (NN) as surrogate models to interpolate TALYS-2.0 outputs based on input nuclear model parameters.

• Data Generation:

  • Datasets were generated using TALYS-2.0 for proton-induced reactions on $^{139}\text{La}$ and $^{\text{nat}}\text{Cu}$.
  • Three levels of complexity were tested: 3-parameter, 6-parameter, and 17-parameter models.
  • Sampling methods compared included Uniform Random, Latin Hypercube (LHC), and Sobol sequence sampling.
  • Hardware utilized included an HPC cluster (BNL) and standalone CPUs (Ryzen 7960X, Intel i7-13700).

• Neural Network Architecture:

  • Framework: PyTorch 2.5.1 with Python 3.11.
  • Structure: A single fully connected deep feedforward NN.
  • Optimization: AdamW optimizer with Huber loss.
  • Regularization: Weight decay, batch normalization, and dropout were used to prevent overfitting.
  • Activation Functions: SELU (input), Mish (hidden), and ReLU (output).
  • Normalization: Cross-sections were normalized by the maximum cross-section of the respective reaction channel.

• Training and Validation Strategy:

  • Data was split into training, validation, and test sets.
  • The model was trained to predict residual product cross-sections (0–200 MeV) given specific nuclear model parameters.
  • Performance was evaluated using Root Mean Squared Error (RMSE), Maximum Absolute Error (MAE), and the 99th percentile absolute error (P99).

• Parameter Adjustment Workflow:

  • 1-D Approach: Sequential optimization of parameters (converging one by one).
  • N-D Approach: Multi-parameter optimization using a dual annealing algorithm for constrained optimization.

3. Key Contributions

• Surrogate Model Implementation: Successfully demonstrated that NNs can serve as high-fidelity surrogates for TALYS-2.0, capable of interpolating complex nuclear reaction outputs across 17 variable parameters.
• Sampling Method Analysis: Showed that Sobol sequence, Latin Hypercube, and Uniform Random sampling yield comparable performance for NN training, with no single method offering a distinct advantage for this specific domain.
• Scalability: Proved the approach scales from simple 3-parameter models to realistic 17-parameter scenarios (matching previous literature on $^{139}\text{La}$) without requiring architectural changes.
• Generalizability: Validated that hyperparameters optimized for $^{139}\text{La}$ transfer effectively to $^{\text{nat}}\text{Cu}$, suggesting the method is applicable to various target nuclei.
• Optimization Efficiency: Introduced a multi-parameter (N-D) adjustment workflow that is significantly faster than traditional sequential methods while achieving better agreement with experimental data.

4. Key Results

• Speedup: The NN approach is >1,000x faster than running TALYS-2.0 directly for generating cross-sections.
  • Generating 100,000+ parameter evaluations took ~8 minutes with the NN (vs. days/weeks with TALYS).
  • A high-fidelity model can be built with ~1,500 TALYS-2.0 files (training time <24 hours on a standard CPU; <few hours on HPC/GPU).
• Accuracy:
  • 3-Parameter Model: The NN outperformed a standard cubic interpolator, especially with smaller datasets (6x6x6 grid), achieving lower RMSE and MAE.
  • 17-Parameter Model: With 2,048 training files (Sobol sampling), the NN achieved an RMSE of 0.75 mb and MAE of 86.7 mb on a 25k-file test set.
  • Sampling Comparison: NNs trained with 1,536 Uniform Random files performed nearly identically to those trained with 2,048 Sobol files.
• Parameter Adjustment:
  • The N-D (multi-parameter) adjustment procedure achieved better agreement with TALYS-2.0 and slightly improved the Figure of Merit (FOM) compared to the 1-D sequential approach and previous literature (Morrell et al.).
  • While some adjusted parameters differed from previous studies (likely due to sensitivity issues or convergence criteria differences), the resulting cross-sections matched experimental data well.
• Robustness: The model successfully handled short-lived parents and complex reaction channels without modification.

5. Significance and Impact

• Parallelization: The workflow decouples the computationally expensive TALYS-2.0 generation (which can be parallelized across HPC clusters) from the parameter fitting (which is instantaneous via the NN). This breaks the bottleneck of sequential processing.
• Feasibility of Global Optimization: The speed of the surrogate model enables the use of advanced global optimization techniques (like dual annealing) and extensive sensitivity analyses that were previously impractical due to time constraints.
• Flexibility: Once trained, the surrogate model allows researchers to change experimental datasets, Figures of Merit, or reaction channels instantly without re-running TALYS-2.0.
• Future Applications: This approach paves the way for rapid isotope production optimization, real-time nuclear data evaluation, and the creation of multi-target nuclear models that require fewer training files than standalone models.

In conclusion, this work establishes that Artificial Neural Networks are a viable, highly efficient, and accurate method for accelerating nuclear reaction simulations, transforming the parameter adjustment process from a time-consuming sequential task into a rapid, parallelizable workflow.