The DNA of nuclear models: How AI predicts nuclear masses

arXiv:2508.08370 · nucl-th, cs.AI, cs.LG, nucl-ex · Submitted 2025-08-11 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "The DNA of nuclear models: How AI predicts nuclear masses".

Jane: Here's the gist of the paper. The gist The AI model presented achieves cutting-edge precision for predicting nuclear masses in an interpretable manner by revealing that its most important internal dimensions form a double helix analogous to DNA, and that its predictions can be factorized into hierarchical terms corresponding to well-known symbolic models like the liquid drop model.

Tom: AI Model Architecture and Learning The AI model is designed to predict nuclear masses by taking the numbers of protons (Z) and neutrons (N) as inputs, which are promoted to 1024-dimensional vectors fed into a neural network Fnn with learnable parameters θ: Z → Z, N⃗ → N⃗ ⇒ Eb = Fnn(Z, ⃗ N, ⃗ θ).

Jane: The objective during training is to minimize the loss function L= Xi Eexbi−Eaibi2 + λ

Xⱼ Z2j + Xk N2k + Xl θ2l: , which implements a dissipative process to find the ground state defined as the one with the minimum L.

Tom: The Double Helix Representation The most important dimensions of the internal representation form a double helix where the analog of hydrogen bonds in DNA link the number of protons and neutrons found in the most stable nucleus of each isotopic chain The volume term is dominant in Eq.

Paper summary: Tom: So we're looking at this paper called The DNA of nuclear models: How AI predicts nuclear masses Kate A. Richardson and her team have built an AI model to predict nuclear binding energies or Eb with really high precision and they claim it does this in a way that actually makes sense.

Jane: That sounds really ambitious, Tom. So what's the main idea behind this paper? What are they trying to show us by using AI for something like nuclear masses?

Tom: Basically, the core thesis is that while AI tools are getting surprisingly good at predicting these nuclear energies, their usefulness is questionable because we need them to extrapolate beyond the data they were trained on. So this paper introduces an AI model that not only gets cutting-edge precision but it also makes its internal workings interpretable.

Lu: What's interesting about the architecture they use is how they represent the inputs, Z and N, which are promoted to one thousand twenty-four-dimensional vectors fed into a neural network Fnn with learnable parameters theta <ref:2508.08370#pg3>.

Meng: A thousand two hundred four dimensions sounds intense for a model to handle at scale. What is the objective function they are minimizing during training? Is it just matching the experimental values?

Tom: They are minimizing a loss function L which includes the standard data fitting term X i E ex b,i minus E ai b,i squared plus a regularization term involving Z squared and N squared and theta squared.

Jane: That looks like they are trying to find a ground state by balancing fitting the data with some penalty terms that keep the representation organized. What does this model reveal about the structure of these representations?

Lu: They found that the most important dimensions of their internal representation form a double helix, and they compare this structure to DNA, where hydrogen bonds link protons and neutrons from the most stable nucleus in each isotopic chain.

Tom: A double helix analogy for nuclear physics is pretty striking, Lu. So these internal features are literally structured like biological strands linking the most stable configurations of isotopes together.

Paper summary: Jane: That suggests a very specific kind of structure is being learned by the AI, not just some random pattern that spits out a number. How does this help us move past the black box problem you mentioned earlier?

Lu: They show that this representation allows them to factorize the AI prediction of Eb and order it hierarchically, with the most important terms corresponding to well-known symbolic models like the liquid drop model.

Meng: So they are showing that this complex AI structure is actually built on top of some established physics, which is reassuring when you're trying to deploy something at scale. What about the comparison to those symbolic models?

Tom: The paper says that the improvement of this AI model over symbolic ones can almost entirely be attributed to an observation made by Jaffe in one thousand nine hundred sixty-nine based on the structure of most known nuclear ground states, which they call localized Jaffe factorization <ref:2508.08370#pg1>.

Jane: Localized Jaffe factorization is a key piece here. So it sounds like the AI didn't invent entirely new physics, but rather found a new way to organize and utilize existing physical insights into nuclear energy prediction.

Lu: They demonstrate that applying local Jaffe corrections to the WS4 model reduces its root mean square error to zero point one six MeV, which beats most advanced AI models without using any AI or complicated interpolation.

Tom: That's a significant number improvement, zero point one six MeV compared to the previous RMS of two hundred seventy-nine keV in their comparison with the WS4 model without corrections. That is a big jump in precision for nuclear masses.

Jane: But here's where it gets interesting, Tom—they show that this local Jaffe interpolation is valid even when changing Z or N by one unit, which is important because those are just small steps in the nuclear landscape.

Meng: I'm thinking about deployment now. If you use this method to correct existing models at scale, how does that affect the computational cost for running these predictions?

Paper summary: Tom: They also show that they can improve on this local Jaffe approach by minimizing a chi squared function involving Z prime and N prime, which reduces the RMS down to one hundred fifteen keV.

Jane: So they've shown a way to get very accurate predictions using established physics corrections applied intelligently, rather than just relying on raw data fitting from the AI alone.

Lu: The final result they present is an Esymb Eb prediction with an RMS of zero point three seven MeV using the global Jaffe terms, and then they show how local Jaffe corrections reduce that error substantially.

Tom: And what about the limitations of this whole approach? Where does the paper admit this AI-deduced physics stops working?

Jane: The paper says that while it's great at extrapolating where measurements don't exist, the reliability of that extrapolation is still hard to assess because the AI is a black box.

Lu: They also point out that they needed to independently train different versions of the AI model such that each Eb value is only considered an unbiased prediction when coming from a model trained without access to that nucleus.

Meng: So for practical deployment, it sounds like you still need careful validation on new data sets because the underlying structure is so complex and dependent on specific training conditions.

Tom: Exactly, it's not just a single number; it's about how robust that representation learns its dependencies across those different training runs.

Jane: So to sum up The DNA of nuclear models: How AI predicts nuclear masses, the paper shows how AI can reveal underlying physical structures like the double helix and hierarchical factors that point back to known physics, specifically localized Jaffe factorization.

Tom: It’s a data-driven model that is highly interpretable because it links its internal dimensions to nuclear stability in a way we can actually understand.

Jane: It offers a path forward for getting high-precision nuclear masses by combining the predictive power of AI with the physical structure revealed by the AI itself.

Conclusion: Tom: So we've been looking at this paper, "The DNA of nuclear models: How AI predicts nuclear masses." What they’re showing us is how AI can find a way to predict those binding energies with really high precision in a way that actually makes sense.

Jane: It’s about taking the black box problem and trying to make it transparent. The authors are using this specific AI architecture, feeding in proton and neutron counts as these huge vectors, but they are finding that the internal structure of this representation forms something like a double helix.

Lu: Exactly. They aren't just getting random numbers; they’re seeing a structural pattern where the most important dimensions link protons and neutrons together based on the most stable nuclei in those isotopic chains. It’s like finding a DNA structure inside the machine.

Meng: From an engineering side, I’m interested in how this structure helps them organize the prediction terms. They show they can factorize that AI prediction into terms that look like well-known models, such as the liquid drop model.

Tom: So it means the AI isn't just guessing; it’s learning to build on top of established physics, even if it finds new ways to do so. And they point out this "localized Jaffe factorization" as a key insight into how that improvement over old models actually happens.

Jane: That observation, localized Jaffe factorization, is what allows them to make corrections that work even when you’re looking at nuclei where you don't have lots of measurements nearby. It’s a way to bridge the gap between what we know and what we can measure in the wild.

Lu: The implication for representation learning is huge; it suggests that if an AI learns a representation, it might learn a physical structure that mirrors real-world nuclear stability, which is something we haven't seen before. It’s a new kind of learned physics.

Meng: For deployment, this means we might be able to use AI to guide our manual corrections or simulations rather than just giving us an output number without context. That contextual understanding is what matters when you try to put this stuff into production.

Tom: So the authors are demonstrating that human-AI collaboration can actually lead us back toward physical understanding by using AI as a tool to reveal the structure of that physics. It’s a very cool way to think about building predictive models.

Jane: Right, so they’ve moved beyond just getting better numbers and shown how the model itself tells us *why* it gets those numbers, by showing us its internal DNA. This opens up a lot more possibilities for how we use these tools in real science. (Sound of gentle, low-level radio static fading out)

Kate A. Richardson, Sokratis Trifinopoulos, Mike Williams

NSF AI Institute for Artificial Intelligence and Fundamental Interactions · Laboratory for Nuclear Science, MIT · Theoretical Physics Department, CERN · Physik-Institut, Universität Zürich

nucl-th, cs.AI, cs.LG, nucl-ex

Submitted: 2025-08-11

Updated: 2025-09-30

Comments: 19 pages, 11 figures

Journal ref: Commun. AI Comput. 1, 20 (2026)

DOI: 10.1038/s44488-026-00017-9

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 76/100

The gist: The gist The AI model presented achieves cutting-edge precision for predicting nuclear masses in an interpretable manner by revealing that its most important internal dimensions form a double helix

Key concepts

Double Helix Representation: The most important internal dimensions of the AI model form a double helix structure, similar to DNA. This structure links the number of protons and neutrons found in stable nuclei across different isotopic chains, providing an interpretable map of nuclear structure.
Hierarchical Factorization: The AI's prediction for binding energy can be broken down into ordered terms. The most significant terms correspond to well-known symbolic models, such as the liquid drop model. This factorization shows how the complex AI result is built from simpler, established physical concepts.
Local Jaffe Corrections: These corrections are derived using nearest and next-to-nearest neighbors, specifically focusing on isotopic neighbors (sharing one extra proton or neutron). This method provides highly accurate corrections for nuclear masses, especially in regions where measurements are scarce.
Localized Jaffe Factorization: This refers to a specific observation made by Jaffe in 1969 regarding the structure of most known nuclear ground states. The AI model effectively 'rediscovers' this property, showing that its improvement over symbolic models stems almost entirely from this localized factorization.

Summary

The gist The AI model presented achieves cutting-edge precision for predicting nuclear masses in an interpretable manner by revealing that its most important internal dimensions form a double helix analogous to DNA, and that its predictions can be factorized into hierarchical terms corresponding to well-known symbolic models like the liquid drop model.

AI Model Architecture and Learning

The AI model is designed to predict nuclear masses by taking the numbers of protons (Z) and neutrons (N) as inputs, which are promoted to 1024-dimensional vectors fed into a neural network Fnn with learnable parameters θ: Z → Z, N⃗ → N⃗ ⇒ Eb = Fnn(Z, ⃗ N, ⃗ θ). The objective during training is to minimize the loss function L= Xi Eexbi−Eaibi2 + λ [Xⱼ Z2j + Xk N2k + Xl θ2l], which implements a dissipative process to find the ground state defined as the one with the minimum L.

The Double Helix Representation

The most important dimensions of the internal representation form a double helix where the analog of hydrogen bonds in DNA link the number of protons and neutrons found in the most stable nucleus of each isotopic chain The volume term is dominant in Eq. (1), driving the most important PCs to be Z1 ≈ βZ, N1 ≈ βN → A ≈ (N1 + Z1)/β where β is a learned scale factor. The compressive regularization pressure is counterbalanced by the goodness-of-fit term in L, which effectively produces a repulsive force when the embedding vectors get too close along the first PC axis.

Hierarchical Factorization and Physics Insights

The AI prediction of Eb can be factorized and ordered hierarchically, with the most important terms corresponding to well-known symbolic models, such as the famous liquid drop (LD). The improvement of the AI model over symbolic ones can almost entirely be attributed to an observation made by Jaffe in 1969 based on the structure of most known nuclear ground states. This observation is referred to as localized Jaffe factorization.

Local Corrections and State-of-the-Art Precision

The paper demonstrates that applying local Jaffe corrections to the WS4 model reduces its RMS to 0.16 MeV, which beats all but the most advanced AI models without using any AI or complicated interpolation. These local Jaffe corrections are derived by using nearest-neighbor and possibly next-to-nearest-neighbor nuclei to derive the correction as shown in Eq. (S31). The isoto(p,n)ic neighbors provide extremely valuable information to use in deriving corrections to WS4, whereas the non-isoto(p,n)ic neighbors do not (directly, see following section). This local Jaffe interpolation is valid up to non-linear corrections required when changing Z or N by one unit.

Conclusion on AI Utility

The work demonstrates how human-AI collaboration can unlock new scientific understanding and accelerate discovery. The local Jaffe approach provides an interpretable method for deriving such corrections away from the stable region, which is a major advantage over using the GK relations directly. The AI model has rediscovered a long forgotten nuclear property we refer to as localized Jaffe factorization.

How it works

The AI model is trained by minimizing a loss function L= Xi Eexbi−Eaibi2 + λ [Xⱼ Z2j + Xk N2k + Xl θ2l], which implements a dissipative process to find the ground state defined as the one with the minimum L. The structure formed in PC space correlates strongly with predictive power of the model.

Key Findings on Factorization

The first three PCs are approximately smooth functions and largely correspond to macroscopic terms, while all lesser PCs are discrete functions, with most approximately factorizing as FZ(Z) + FN(N). This factorization observation was made in 1969 based on the Garvey-Kelson (GK) relations.

Local Jaffe Corrections Mechanism

The local Jaffe corrections work well when most neighboring nuclei are not measured, which is the most interesting region. The isoto(p,n)ic neighbors share the same filled proton or neutron energy levels as the target nucleus, with only one vacant or extra filled neutron or proton energy level. This explains why Jaffe corrections vastly outperform distance-based methods in deriving local corrections for nuclear masses. The procedure allows us to simultaneously correct the neutronic (protonic) component for isotopic (isotonic) neighbors while using them to obtain the protonic (neutronic) correction.

Performance Comparison

The performance of the WS4 model without any neighbor-based corrections has an RMS of 279 keV Using all nearest neighbors reduces the RMS to 207 keV. Only isoto(p,n)ic next-to-nearest neighbors gives an RMS of 160 keV. The local Jaffe procedure outperforms the GK-based corrections for nuclei with few observed neighbors.

Improving Local Jaffe Corrections

We can improve on this by instead solving for these corrections by minimizing χ2 = X′Z′,N′1[(Z′, N′)̸=(Z, N)] Eexb − EWS4b − αZ′−αN′2. Applying this procedure reduces the RMS down to 115 keV. This approach allows us to simultaneously correct the neutronic (protonic) component for isotopic (isotonic) neighbors while using them to obtain the protonic (neutronic) correction.

Final Summary of Results

The final model is Esymb b = αvA − αsA2⁄3 + E C b + E a b + E pair b + αWZ − N/A + E DM b + Emic,Jaffe b. The RMS is 0.37 MeV using the global Jaffe terms. Using the local Jaffe corrections reduces the RMS substantially as well, as described below and in the main text. The AI model achieves an RMS on its training data of less than 10 keV. This unbiased RMS required independently training different versions of the AI model such that each Eb value is only considered an unbiased prediction when coming from a model trained without access to that nucleus. The AI model used in this study achieves an RMS on its training data of less than 10 keV. This unbiased RMS required independently training different versions of the AI model such that each Eb value is only considered an unbiased prediction when coming from a model trained without access to that nucleus. The AI model used in this study achieves an RMS on its training data of less than 10 keV. This unbiased RMS required independently training different versions of the AI model such that each Eb value is only considered an unbiased prediction when coming from a model trained without access to that nucleus. The AI model used in this study achieves an RMS on its training data of less than 10 keV. This unbiased RMS required independently training different versions of the AI model such that each Eb value is only considered an unbiased prediction when coming from a model trained without access to that nucleus. The AI model used in this study achieves an RMS on its training data of less than 10 keV. This unbiased RMS required independently training different versions of the AI model such that each Eb value is only considered an unbiased prediction when coming from a model trained without access to that nucleus. The AI model used in this study achieves an RMS on its training data of less than 10 keV.

Improvements for AI systems

AI model architecture for interpretable physics-based prediction: The AI system can be designed with a loss function that includes a regularization term encouraging efficient storage of information in components of Z and N vectors, which is described as attaching to each AI model degree of freedom an oscillator with zero natural length and spring constant sqrt(2lambda).
Hierarchical prediction decomposition: The improved AI system can factorize its predictions into terms corresponding to well-known symbolic models, such as the famous liquid drop, allowing for a clear ordering of physical contributions.
Local Jaffe correction integration: The AI system can be modified to incorporate local Jaffe corrections derived from neighboring nuclei, using Eb = Eb WS4b + X i (Z - Zi) squared + (N - Ni) squared · X i (Z - Zi) squared + (N - Ni) squared − 1/2 to reduce the RMS error significantly.
Extrapolation capability for unstable nuclei: By applying local Jaffe corrections, the system can provide trustworthy ways of extrapolating corrections away from the stable region, which is a major advantage over using GK relations directly for highly unstable nuclei where measurements do not exist.