From alpha decay to cluster decay: an extreme case of transfer learning

arXiv:2603.28628 · nucl-th · Submitted 2026-03-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Nuclear Theory. The day in nuclear theory: structure, reactions and the models behind them.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "From alpha decay to cluster decay".

Paper summary: Kai: So wrapping up this talk on "From alpha decay to cluster decay: an extreme case of transfer learning", we’re seeing how transfer learning can bridge the gap between data-rich and data-scarce problems in nuclear physics.

Mira: The authors are using alpha decay half-lives as a source domain to inform a model that predicts cluster decay half-lives in the target domain, showing that this reuse of knowledge stabilizes optimization when training data is limited.

Lev: What this really means for us on the hardware side is that we might be able to develop more robust models for nuclear structure predictions even when we only have limited experimental data for the specific phenomena we care about.

Kai: The paper’s conclusion points toward this transfer learning method as a way to achieve stable and reliable learning in the small-sample regime where deep neural networks usually struggle with random initialization and sampling bias.

Mira: They've demonstrated that even with modest datasets, if you start with physically informed pretraining on something abundant like alpha decay, you can effectively adapt it to predict cluster decays using only a small cluster decay dataset for fine-tuning.

Lev: It suggests that the physical commonality between these two types of nuclear decay provides a strong enough regularizer to make deep learning work in these challenging areas.

Kai: The authors are showing that this transfer learning approach is an extreme case where it works well, even when the data scarcity is quite severe.

Mira: It shifts the focus from needing massive datasets for every specific decay mode to finding effective ways to share physics between related decay modes through transfer learning architectures.

Conclusion: Kai: So, we just looked at how they use alpha decay data to teach a model for cluster decay, and now we need to talk about what this whole paper is actually saying about that transfer learning idea.

Mira: The title itself tells you the core idea—it's taking knowledge from one type of nuclear process, alpha decay, and using it to predict another, cluster decay. They're calling it an extreme case because they’re doing this when the data for the second thing is pretty sparse.

Lev: From my side, what I see is that they are showing how you can keep your model from just getting totally confused when you don't have tons of examples for the new process. The pretraining on alpha decay gives it a solid foundation before they even look at the cluster decay data.

Kai: Yeah, and Mira’s right about the title—it’s not some casual transfer; it’s a deliberate setup where one dataset is huge and informative, and we use that to make sense of a much smaller dataset for something else. It really boils down to how you structure the learning process when you're fighting data scarcity.

Mira: And what this means for us as theorists is that maybe we don't need a massive, dedicated cluster decay database right away if we can find a similar decay mode, like alpha decay, and use that physics to kickstart our predictions. It suggests physics has common ground across these nuclear structures.

Lev: For the hardware guys, it means when we build an error correction scheme for something new and data-poor, having a model pre-trained on something abundant can make the whole thing more stable from the start. It helps avoid those bad initial conditions that usually mess up deep learning when you only have a few samples.

Kai: Exactly. The authors are showing that even in this extreme setup, you can get reliable results by smartly sharing physical knowledge between these related decay modes. It shifts our focus from just needing more raw data for every single thing to finding the right ways to share physics between different nuclear decay channels and models.

Mira: So, the authors are essentially arguing that transfer learning isn't just a neat trick; it’s a practical tool for dealing with the real-world limitations of having limited experimental data in complex fields like nuclear structure. We have to figure out how to leverage what we already know.

Sino-French Institute of Nuclear Engineering and Technology · Sun Yat-sen University

nucl-th

Submitted: 2026-03-30

Updated: 2026-11-17

Comments: 9 pages, 4 figures

Journal ref: Phys. Rev. C 114, 014314(2026)

DOI: 10.1103/r796-y5tc

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 79/100

The gist: From α decay to cluster decay: an extreme case of transfer learning addresses data scarcity in nuclear physics by pretraining a deep neural network on abundant alpha decay data and fine-tuning it on…

Key concepts

Alpha Decay
Alpha decay is a nuclear process where an unstable nucleus spontaneously emits an alpha particle (two protons and two neutrons). The paper uses this abundant data as the source domain for pretraining. This large dataset provides the deep neural network with foundational knowledge about nuclear structure and decay mechanisms.
Cluster Decay
Cluster decay refers to a different type of nuclear disintegration where the nucleus breaks into two or more smaller, stable sub-nuclei called clusters. This process is less frequently observed in experiments, making data for it scarce. The goal is to accurately predict these rare events using the knowledge gained from alpha decay.
Transfer Learning
Transfer learning is a machine learning technique where a model trained on one task (the source task, like alpha decay) is repurposed and adapted for a different, related task (the target task, like cluster decay). Instead of training from scratch on the small cluster data, the pre-trained weights are used as an excellent starting point.
Deep Neural Network
A deep neural network is a complex type of artificial intelligence model with many layers designed to learn intricate patterns in data. In this context, it learns the relationship between nuclear properties and decay types. The paper uses its depth to capture the complex physics embedded within the input data for both alpha decay and cluster decay predictions.

Terminology

Summary

From α decay to cluster decay: an extreme case of transfer learning addresses data scarcity in nuclear physics by pretraining a deep neural network on abundant alpha decay data and fine-tuning it on scarce cluster decay data to achieve accurate predictions p. 0

Where the paper says it

From α decay to cluster decay: an extreme case of transfer learning Yinu Zhang,∗ Zhiyi Li, Kele Li, Jiaxuan Zhong, and Cenxi Yuan† Sino-French Institute of Nuclear Engineering and Technology, Sun Yat-Sen University, Zhuhai 519082, China (Dated: March 31, 2026) When training data are limited, data-driven models are especially vulnerable to optimizationrelated fluctuations from random initialization and to sampling-induced bias from insufficient training data. We address both challenges with transfer learning (TL): deep neural networks (DNNs) are first pretrained on α decay half-lives and then fine-tuned on a small cluster decay dataset.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Transfer learning can mitigate both optimization-related fluctuations from random initialization and sampling-induced bias from insufficient training data. This allows deep neural networks to be pretrained on a rich source domain, like alpha decay half-lives, providing a physically informed initialization that stabilizes optimization.

  2. The improved AI system can accurately predict cluster decay half-lives for parent nuclei from 221Fr to 242Cm despite the extreme data scarcity of cluster decay experimental data. This is achieved by leveraging the learned global decay systematics from the pretraining stage to regularize the fit and reduce sensitivity to training set composition.

  3. The system can achieve stable and reliable learning in small-sample regimes by using transfer learning, as demonstrated when a model initialized with pretrained parameters performs better than one trained directly on limited cluster decay data. This reduces the instability caused by nonconvex nature of DNN optimization when training on very small target datasets.

  4. The AI system can be optimized to use two distinct fine-tuning regimes: full fine-tuning and shallow fine-tuning for cluster decay prediction. Full fine-tuning reoptimizes all weights, while shallow fine-tuning restricts updates to the last few layers, allowing the model to leverage shared physics while limiting overfitting in the target domain.

  5. The AI system can utilize a physically informed inductive bias during fine-tuning by replacing random initialization with a physically informed starting point learned from the much larger alpha decay dataset. This constrained initialization leads to more stable predictions and reduces optimization-related fluctuations compared to training from arbitrary random weights.

Abstract

When training data are limited, data-driven models are especially vulnerable to optimization-related fluctuations from random initialization and to sampling-induced bias from insufficient training data. We address both challenges with transfer learning (TL): deep neural networks (DNNs) are first pretrained on α decay half-lives and then fine-tuned on a small cluster decay dataset. The pretraining stage provides a physically informed initialization that stabilizes optimization, while transferred global decay systematics regularize the fit and reduce sensitivity to training set composition. Despite extreme data sparsity, the resulting models accurately predict cluster decay half-lives for parent nuclei from 221 Fr to 242 Cm. We further quantify how initialization and sample selection affect predictive accuracy and robustness, demonstrating that TL enables stable and reliable learning in the small-sample regime.

Sources

Related papers