betaFeedback

WorksPhysics and Astronomy

Global Framework for Emulation of Nuclear Calculations

Abstract

We introduce a hierarchical framework that combines ab initio many-body calculations with a Bayesian neural network, developing emulators capable of accurately predicting nuclear properties across isotopic chains simultaneously and being applicable to different regions of the nuclear chart. We benchmark our developments using the oxygen isotopic chain, achieving accurate results for ground-state energies and nuclear charge radii, while providing robust uncertainty quantification. Our framework enables global sensitivity analysis of nuclear binding energies and charge radii with respect to the low-energy constants that describe the nuclear force.

Introduction. — The study of nuclear properties and the interactions that govern them is critical to understanding how complex many-body systems such as nuclei, atoms, and molecules emerge from the fundamental forces and particles of nature. At the most fundamental level, the nuclear force is described by Quantum Chromodynamics (QCD), the theory of strong interactions. Due to the non-perturbative complexity of QCD in the low-energy regime, effective interactions derived from Chiral Effective Field Theory (χ\chiEFT) [1, 2] are commonly employed to describe the interactions between neutrons and protons inside the atomic nucleus. χ\chiEFT provides a low-energy expansion of QCD in terms of nucleon-nucleon and three-nucleon interactions, with undetermined Low-Energy Constants (LECs) adjusted to reproduce few-nucleon observables. However, nuclear properties can be highly sensitive to the value of these LECs [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. Therefore, it is crucial to understand the impact of the LECs on nuclear observables and how their uncertainties propagate through ab initio many-body calculations.

Ab initio methods [18, 19, 20, 21, 22, 23, 24, 25] rely on systematically improvable approximations to solve the many-body Schrödinger equation, starting from χ\chiEFT interactions. Despite major advances in recent years, ab initio nuclear calculations of medium and heavy nuclei remain prohibitively expensive due to the exponential scaling of many-body wave functions. While approximate polynomially-scaling methods (e.g. [19, 20]) have been developed, this complexity still imposes stringent resource demands and methodological challenges [26, 27, 28]. Statistical procedures, such as global sensitivity studies, are necessary to evaluate nuclear observables for samples in the range of millions, even for the simplest realistic nuclear interactions [29]. As a result, surveying the entire nuclear landscape with realistic uncertainty estimation and establishing a connection to the LECs remains an outstanding challenge.

Nuclear emulators have emerged as powerful tools to address these problems, offering efficient approximations of many-body calculations at substantially reduced computational cost and allowing for uncertainty quantification [30]. Techniques such as multi-output multi-fidelity Deep Gaussian Processes [17, 31], eigenvector continuation [29, 32, 33, 34, 35] and many-more (e.g. [36, 37, 38]) have been proposed as surrogate models that approximate the outputs of nuclear calculations with significantly reduced computational costs. However, these methods are so far limited to individual isotopes, restricting their ability to capture correlations and trends across isotopic chains and more globally. Moreover, they still demand costly ab initio calculation train them.

To overcome the challenges mentioned above, we introduce BAyesian Neural Network for Atomic Nuclei Emulation (BANNANE), a hierarchical Bayesian neural network (BNN) framework that integrates multi-fidelity datasets with categorical and positional embeddings to emulate nuclear properties globally. This approach introduces a flexible architecture tailored for the prediction of different nuclear properties with vastly reduced computational time. We demonstrate the overarching capabilities of BANNANE by calculating the nuclear binding energies and charge radii of the oxygen isotopic chain, achieving accurate predictions of these observables. Notably, BANNANE enables a global analysis of the sensitivity of these observables to the LECs used in the interactions. This allows us to directly study how macroscopic nuclear properties are impacted by specific details of the nuclear force, and how these sensitivities vary along isotopic chains, which, to our knowledge, is not possible with current nuclear emulators.

Methods. — We construct a multi-fidelity Bayesian emulator using a hierarchical architecture tailored for global nuclear modeling. Our pipeline is divided into three main stages: (i) data loading and preprocessing, (ii) hierarchical BNN construction, and (iii) variational training and inference.

Data generation and preprocessing: LEC samples utilized in this work are taken from Ref. [35], where history matching was used to constrain the LECs to physically plausible ones. More specifically, these samples are for a formulation of χ\chiEFT where Δ\Delta-isobars are considered explicitly, up to next-to-next-to leading order (N2LO) in the chiral expansion. In particular, 17 LECs are required to parametrize this theory. We construct data sets for training, validation, and testing randomly from the 8188 samples given in Ref. [35]. We employ the Valence-Space formulation of the In-Medium Similarity Renormalization Group (VS-IMSRG) [20, 39] to solve the many-body problem and obtain the ground-state energies and charge radii of the different isotopes. In practice, the decoupling of the valence space is done using the imsrg++ code [40], and the valence space is then diagonalized using the KSHELL shell-model code [41]. In particular, we use calculations with different model space sizes, given by including all harmonic oscillator states with e=2n+l≤emax⁡e = 2n + l \le e_{\max} where nn and ll are the principal quantum number and angular momentum of the wave functions respectively. We note that increasing emax⁡e_{\max} results in a more accurate calculation, at the cost of a rapid rise in computational power required to solve the problem. We therefore consider the values of the observables at emax⁡∈{4,6,8,10}e_{\max} \in\{4, 6, 8, 10\} to constitute different fidelities in the context of the emulator.

Hierarchical BNN: Figure 1 presents a diagrammatic representation of the multi-fidelity architecture used in BANNANE. Our core model uses distinct embeddings for the proton number ZZ, and neutron number NN, via a learnable positional encoding, and emax⁡e_{\max} of each sample. These embeddings are concatenated with the LECs, then passed through a shared multi-head self-attention layer that captures correlations across nuclei and is queried for each fidelity. To accommodate multiple fidelities, we combine a base predictor (representing the lowest-emax⁡e_{\max} dataset) with additive delta blocks for each higher-emax⁡e_{\max} level. Concretely, the network first produces a common latent representation, which is mapped to predictions via (i) a base model for the lowest emax⁡e_{\max} and (ii) incrementally learned offsets for any higher-emax⁡e_{\max} data present. This ensures consistent cross-emax⁡e_{\max} learning, allowing the emulator to leverage coarse information from lower-emax⁡e_{\max} calculations and refine it where higher-emax⁡e_{\max} data exist.

BANNANE architecture overview

Figure 1. BANNANE architecture overview. The input LECs, along with categorical embeddings for proton number (ZZ), neutron number (NN), and fidelity level (emax⁡e_{\max}), are processed through a Bayesian shared-latent layer and a multi-head attention mechanism. Fidelity-specific prediction heads refine the base prediction at higher emax⁡e_{\max} for each fidelity ranging from 2, outputting mean (μNf\mu_N^f) and uncertainty (σNf\sigma_N^f) estimates for nuclear observables.

Variational training and inference: We implement the hierarchical BNN with Pyro’s stochastic variational inference (SVI) [42], employing a diagonal Gaussian variational posterior. All weights, biases, and emax⁡e_{\max}-specific parameters (including output variance terms) are learned by minimizing evidence lower bound (ELBO) [43]. After each SVI iteration, the validation loss is monitored for early stopping. During inference, our posterior predictive distribution is sampled for each (Z,N)(Z,N) and emax⁡e_{\max}. This yields both mean predictions and standard deviations, thus quantifying uncertainties across isotopic chains. By design, BANNANE can extrapolate to new or sparse regions of the nuclear chart simply by evaluating the learned embeddings of (Z,N)(Z,N). Further implementation details and additional ablation studies are provided in the supplemental material [44].

Results.— In this section, we demonstrate BANNANE’s performance under a range of conditions:

Performance on the oxygen isotopic chain: We assess BANNANE on the oxygen isotopic chain (12{}^{12}O–24{}^{24}O) including up to emax⁡=10e_{\max}=10 for all the isotopes. Figure 2 displays the predicted ground-state energies at emax⁡=10e_{\max}=10 for all oxygen isotopes, compared to the reference IMSRG calculations.

Full-chain training on the oxygen isotopes

Figure 2. Full-chain training on the oxygen isotopes. BANNANE predictions versus IMSRG reference results for test samples at the highest fidelity emax⁡=10e_{\max}=10, for EBE_B and RchR_{\rm ch}. Error bars indicate 1σ1\sigma uncertainty from BANNANE’s posterior, and the average Root Mean Squared Error (RMSE) is displayed for each one.

Learnt embedding space

Figure 3. Learnt embedding space. Projection of the attention map to 2 dimensions using t-distributed Stochastic Neighbor Embedding (t-SNE) dimensionality reduction applied to the latent space for test LEC samples. The line and colored background show the decision boundary of a simple linear classifier to the reduced space to illustrate the separation between shells.

We observe that BANNANE correctly reproduces both total binding energies EBE_B and charge radii RchR_{\rm ch} across the chain with 0.8 MeV and 0.01 fm Root Mean Squared Error (RMSE), respectively. This represents a significant improvement compared to the Eigenvector Continuation-based subspace-projected coupled-cluster [29] which obtains RMSEs of 3 MeV and 0.02 fm for the ground state energies and charge radii respectively in 16O^{16}\mathrm{O} using 128 training points. While BANNANE uses more training points in total for this isotope (715 in total), only 51 samples are of the highest fidelity, making BANNANE much more computationally efficient to train, while still achieving an RMSE of 0.337 MeV and 0.02 fm, respectively, for the same isotope. Further study of the residuals and benchmark is provided in the supplemental material [44].

Moreover, we note that the discontinuity caused by the shell closure at N=8N=8 is well captured by BANNANE. To further study this phenomenon, we performed a dimensional reduction to perform an analysis of the latent space of the model after the attention mechanism using a projection t-distributed Stochastic Neighbor Embedding (t-SNE) dimensionality reduction [45] for visualization. As illustrated in Fig. 3, there is an evident clustering of the isotopes in the sdsd shell, after the shell closure at N=8N=8, whereas the isotopes in the pp shell appear to be mapped to distinct clusters, hinting of the model’s ability to capture an underlying nuclear structure.

Zero-shot extrapolation: A major advantage of BANNANE is its ability to make reliable predictions for isotopes beyond those included in the training. BANNANE’s architecture enables true zero-shot generalization, meaning it can infer nuclear properties of isotopes it has never seen before, solely from learned trends across isotopic chains. To illustrate this, we simultaneously withheld all the data corresponding to 15O^{15}\mathrm{O} samples and evaluated the model’s regression.

Figure 4 (top) shows the relative errors (as a percentage) of EBE_B once all the samples for this isotope are left out of the training set. Interestingly, the inductive biases on the hierarchical architecture are such that the model can capture both convergence and nuclear structure even for unseen isotopes, as shown by the small residuals at each fidelity. Besides the noticeable shift in the regressed values, it is still surprising that no samples for this isotope were provided. We note that the performance is weaker for the charge radii, especially near the shell closure as discussed along with further analysis for extrapolation over the complete chain in the supplemental material [44].

This fit can be improved significantly when including data only from the lower fidelity, namely from emax⁡=4e_{\max}=4, as illustrated in Figure 4 (bottom). We find that including the lower fidelity greatly reduces the spread of the residuals, reduces the systematic bias, and makes the distribution closer to a Gaussian, making for much better-behaved residuals on all fronts.

This extrapolation feature shows great advantages for future applications to heavy systems where highprecision calculations of many LEC samples are infeasible for an entire isotopic chain. In practice, one could choose to compute emax⁡=10e_{\max}=10 results only for a small strategic set of isotopes (e.g., a few in mid-shell regions where higher-order correlations are most relevant and around shell-closure where discontinuity are more likely to happen) and low-fidelity data for the rest of the isotopes.

Residual distributions for extrapolation to $^{15}\mathrm{O}$

Figure 4. Extrapolation to 15O^{15}\mathrm{O}. (a) Residual distribution of the binding energy EBE_B (%) for BANNANE predictions compared to IMSRG reference values at emax⁡=6,8,10e_{\max}=6, 8, 10, when no training samples for 15O^{15}\mathrm{O} were included. (b) Residual distribution for EBE_B predictions when only low-fidelity data (emax⁡=4e_{\max}=4) was used in training. The inclusion of low-fidelity data significantly reduces systematic bias and improves the overall accuracy of the extrapolated predictions.

Crucially, the Bayesian nature of BANNANE captures the remaining uncertainties due to limited high-fidelity coverage. These uncertainty estimates can inform experimental proposals or guide more selective future high-fidelity calculations, focusing effort on regions where the emulator’s confidence is lowest. This phenomenon can be further studied by analyzing how the model’s emulation residual converges as more high-fidelity training data is included, for which results are presented in the Result section of the supplemental material [44].

Physical Convergence: BANNANE not only interpolates and extrapolates nuclear observables but also emulates the physical convergence of many-body methods. Figure 5 shows the emulator’s predictions for EBE_B across the oxygen chain as a function of emax⁡e_{\max}, compared to experimental values after doing a weighted resampling of 8188 LEC samples [35] as done in Ref. [17]. The emulator accurately reproduces convergence trends, including nontrivial shell effects, with deviations well within its uncertainty estimates, and in good agreement with previous results [48]. Notably, these uncertainties stem solely from the emulator and the LECs, not from many-body truncation effects. This highlights BANNANE’s ability to

Emulator-driven convergence of binding energies for oxygen isotopes

Figure 5. Emulator-driven convergence of binding energies EBE_B for oxygen isotopes as a function of emax⁡e_{\max}. Results are compared to fully converged VS-IMSRG results [46] at emax⁡=14e_{\max}=14 using the EM(1.8/2.0) nuclear interaction in purple squares. Solid lines represent experimental results from [47].

capture systematic trends while reducing computational costs, which in practice could take several years on HPC clusters to just a few seconds for the resampling.

Unified Sensitivity Estimation: Beyond providing fast and accurate emulation, BANNANE enables global sensitivity analysis (GSA) that would be computationally prohibitive using brute-force ab initio methods. With this method, we can probe how much the variance of each LEC (and their covariance) impacts the variance of the observables. In Figure 6, we apply a variance-based Sobol approach [49] to assess how each LEC affects EBE_B and RchR_{\mathrm{ch}} across oxygen isotopes from 12O^{12}\mathrm{O} to 24O^{24}\mathrm{O}. We plot the main-effect index SiS_i (filled bars), capturing the direct contribution of each LEC to the variance of an observable, and the total-effect index STS_T (white outlines), accounting for higher-order interactions (see supplemental material [44] for more details.).

Variance-based global sensitivity analysis for the oxygen isotopes

Figure 6. Variance-based global sensitivity analysis for the oxygen isotopes. Shown are the main-effect indices SiS_i (colored bars) and total-effect indices STS_T (black outline) for each LEC, indicating their direct and combined contributions to the variance of the binding energy EBE_B (top) and charge radius RchR_{\rm ch} (bottom). Notable differences between SiS_i and STS_T point to strong interactions among particular LECs.

A few features stand out: For binding energies (top panel), certain two-nucleon (2N) couplings (e.g., C1S0C_{1S0} and C2C_2) and three-nucleon (3N) couplings (e.g., CDC_D, CEC_E) dominate the variance while showing small progressive changes across the chain. This is consistent with findings of previous methods [15, 17, 29] in multiple nuclei. It indicates that the dependency of the ground state energy to the LEC exhibits small variations in different nuclear systems. In contrast, the hierarchy of LEC sensitivities differs considerably for the charge radii (bottom panel): large changes can be observed at the shell-closure, with for example C3P2C_{3P2} being important at CE1C_{E1} being negligible in the sd-shell but the opposite holding true in the p-shell. Moreover, the sensitivities for RchR_{\mathrm{ch}} are inherently more nonlinear and show the emerging importance of cross-couplings rather than individual influences. This highlights that the charge radius can be highly sensitive to certain LECs, making it strongly complementary to the experimental constraints derived from binding energies. In particular, experimental data on neutron-deficient oxygen isotopes, which have not yet been measured, will be crucial for guiding these developments and benchmarking our predictions.

Conclusions

Conclusions.— We have introduced a hierarchical multi-fidelity Bayesian Neural Network for Atomic Nuclei Emulation (BANNANE), which synthesizes low- and high-fidelity datasets across different nuclei. By leveraging learnable embeddings for nucleon numbers and an additive fidelity-specific architecture, BANNANE not only interpolates faithfully among known nuclei, achieving better accuracy than current emulators but also extrapolates with remarkable accuracy to hitherto uncalculated (or unobserved) nuclei. Due to this, BANNANE offers a computationally feasible way to calculate nuclear properties with realistic uncertainties and with an accuracy comparable to those obtained by considerably costly ab initio methods. These developments have enabled us to investigate the overall dependence of nuclear observables on the LECs that govern inter-nucleon interactions. We found that the nuclear binding energy and charge radii of oxygen isotopes exhibit distinct sensitivities to these LECs, with particular interest in data from neutron-deficient oxygen isotopes that have yet to be measured. We hope our findings will motivate charge radius measurements of these isotopes, which are now feasible via laser spectroscopy experiments [50] at the new Facility for Rare Isotope Beams (FRIB) [51]. More broadly, we anticipate that BANNANE will serve as a powerful tool to guide future experiments and establish direct connections between measurements of binding energy, charge radii, and specific components of the nuclear force.

Beyond its computational advantages, BANNANE provides a powerful framework to guide future experimental and theoretical efforts. It enables targeted high-precision calculations and experimental searches where they matter most, optimizing resource allocation for both theory and experiments. Its versatile architecture can be used with any many-body method and can include as many observables as desired. While we have only varied the neutron number in this study, BANNANE architecture also allows the emulation of different proton numbers, which we will test in future work.

Data Availability

Data Availability.— The data for reproducing the results in this work can be found online [52] along with scripts used in the data analysis. Additionally, the source code for training the model is available upon request.

We thank A. Ekström, C. Forssén, G. Hagen, and W. G. Jiang for providing the interaction samples used in this work, and J. Holt for insightful discussions. The IMSRG code used in this work makes use of the Armadillo C++ library [53, 54]. Computational resources were provided by the SUMMIT at MIT Physics. This work was supported by the Office of Nuclear Physics, U.S. Department of Energy, under grants DESC0021176 and DE-SC0021179. We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC) [PDF-587464-2024].

Methods

Architecture Details

BANNANE is designed to emulate nuclear properties across isotopic chains while integrating multi-fidelity data. Its hierarchical Bayesian framework combines learnable embeddings, a multi-head self-attention mechanism with fidelity-specific queries, and fidelity-specific prediction pathways. This architecture ensures that the emulator can leverage correlations between nuclear observables across elements and isotopes, while also adapting to varying levels of computational fidelity. It must be added that we expect the model to further improve the performance once a full hyperparameter optimization is performed. Below we present a detailed overview of our implementation.

Input embeddings.— To incorporate discrete nuclear identifiers and fidelity information, BANNANE utilizes three primary embedding layers:

Element (ZZ) embedding: An embedding layer maps each proton number Z∈{0,1,…,Zmax⁡}Z \in\{0, 1, \ldots, Z_{\max}\} to a continuous vector space:

eZ=WZ⋅1Z,\mathbf{e}_Z = \mathbf{W}_Z \cdot\mathbf{1}_Z,

where WZ∈R(Zmax⁡+1)×dZ\mathbf{W}_Z \in\mathbb{R}^{(Z_{\max}+1)\times d_Z} is a learnable weight matrix, 1Z\mathbf{1}_Z is a one-hot encoding of ZZ, and dZd_Z is the embedding dimension. To increase the architecture’s flexibility, we left the option to include this embedding into the model, which is not considered for the benchmarks presented in this work but will be needed when other ZZ values are considered.

Neutron number (NN) positional encoding: A similar embedding can be applied to the neutron number NN:

eN=WN⋅1N\mathbf{e}_N = \mathbf{W}_N \cdot\mathbf{1}_N

with WN∈R(Nmax⁡+1)×dN\mathbf{W}_N \in\mathbb{R}^{(N_{\max}+1)\times d_N} and dNd_N the corresponding dimension. This learned positional encoding allows the network to capture trends as a function of neutron number, crucial for modeling isotopic evolution.

However, instead of a learnable embedding, BANNANE uses a fixed sinusoidal positional encoding inspired by Transformer architectures. This encoding defines a phase:

φ=⋅exp⁡(−2idNlog⁡(10000)),i=0,…,⌊dN−22⌋\varphi= \cdot\exp\left(-\frac{2i}{d_N}\log(10000)\right), \quad i=0,\ldots,\left\lfloor\frac{d_N-2}{2}\right\rfloor

Which is then added as a positional encoding via:

PE⁡(n,2i)=sin⁡(nφ),PE⁡(n,2i+1)=cos⁡(nφ).\operatorname{PE}(n,2i)=\sin(n\varphi), \quad\operatorname{PE}(n,2i+1)=\cos(n\varphi).

This guarantees that the neutron number encoding captures smooth trends across isotopic chains, aiding generalization. The sinusoidal encoding is implemented as a configurable alternative to learned embeddings, selectable via hyperparameters. We used the sinusoidal encoded NN embeddings for all the results in this work.

Fidelity level embedding: Each fidelity level f∈Ff \in\mathcal{F} is similarly embedded:

ef=Wf⋅1f,\mathbf{e}_f = \mathbf{W}_f \cdot\mathbf{1}_f,

where Wf∈R∣F∣×df\mathbf{W}_f \in\mathbb{R}^{|\mathcal{F}|\times d_f} maps a one-hot representation of fidelity level to an embedding space of dimension dfd_f.

The continuous input features x∈Rdx\mathbf{x} \in\mathbb{R}^{d_x}, which may include coupling constants and other parameters derived from effective field theories, are concatenated with these embeddings to form a combined input:

z=[x,eZ,eN,ef]∈Rdx+dZ+dN+df.\mathbf{z} = \left[\mathbf{x},\mathbf{e}_Z,\mathbf{e}_N,\mathbf{e}_f\right] \in\mathbb{R}^{d_x+d_Z+d_N+d_f}.

Shared latent representation.— The combined input z\mathbf{z} is fed into a shared latent network to extract a feature-rich representation:

h=ϕ(W2⋅ϕ(W1⋅z+b1)+b2),\mathbf{h} = \phi\left(\mathbf{W}_2 \cdot\phi(\mathbf{W}_1 \cdot\mathbf{z}+\mathbf{b}_1)+\mathbf{b}_2\right),

where W1\mathbf{W}_1, W2\mathbf{W}_2 are weight matrices, b1\mathbf{b}_1, b2\mathbf{b}_2 are biases, and ϕ(⋅)\phi(\cdot) denotes a non-linear activation function (e.g., LeakyReLU). This produces a latent vector h∈Rdh\mathbf{h} \in\mathbb{R}^{d_h} shared across all fidelities.

Multi-head self-attention mechanism with fidelity-specific queries.— To capture complex interdependencies and fidelity-specific features, BANNANE employs a multi-head self-attention mechanism augmented with fidelity-specific query vectors [55]. This allows the model to extract different aspects of the shared representation tailored to each fidelity level. For our implementation, we fixed the number of attention heads to 2.

Fidelity-specific queries: For each fidelity level, including the base fidelity, a unique learnable query vector is introduced:

Q(f)=Wq(f)⋅h,\mathbf{Q}^{(f)}=\mathbf{W}^{(f)}_q\cdot\mathbf{h},

where Wq(f)\mathbf{W}^{(f)}_q is a learnable projection matrix specific to fidelity ff.

Attention computation: For a given fidelity ff, the query Q(f)\mathbf{Q}^{(f)} interacts with the shared latent representation h\mathbf{h} to compute attention weights and extract relevant features:

Attention⁡(f)(Q(f),K,V)=softmax⁡(Q(f)KTdk)V,\operatorname{Attention}^{(f)}(\mathbf{Q}^{(f)},\mathbf{K},\mathbf{V})=\operatorname{softmax}\left(\frac{\mathbf{Q}^{(f)}\mathbf{K}^{\mathsf{T}}}{\sqrt{d_k}}\right)\mathbf{V},

where K=Wkh\mathbf{K}=\mathbf{W}_k\mathbf{h} and V=Wvh\mathbf{V}=\mathbf{W}_v\mathbf{h} are shared key and value projections, respectively, and dkd_k is the dimension per head.

Concatenation and projection: The outputs from all attention heads are concatenated and passed through a linear layer to form the final attention-enhanced representation:

o(f)=Wo[head⁡1(f),…,head⁡H(f)],\mathbf{o}^{(f)} = W_o\left[\operatorname{head}_1^{(f)},\ldots,\operatorname{head}_H^{(f)}\right],

with Wo∈R(H⋅dk)×doutW_o \in\mathbb{R}^{(H \cdot d_k) \times d_{\mathrm{out}}}.

Fidelity-specific regression pathways.— The attention-enhanced representation o(f)\mathbf{o}^{(f)} is used to predict target observables hierarchically, leveraging both shared and fidelity-specific information:

Base prediction: For the lowest fidelity level f0f_0, the model computes

y(f0)=fbase(o(f0))=Wbase2⋅ϕ(Wbase1⋅o(f0)+bbase,1)+bbase,2,\begin{aligned} \mathbf{y}^{(f_0)} = f_{\mathrm{base}}\left(\mathbf{o}^{(f_0)}\right) \\ &= W_{\mathrm{base}}^{2} \cdot\phi\left(W_{\mathrm{base}}^{1} \cdot\mathbf{o}^{(f_0)} + \mathbf{b}_{\mathrm{base},1}\right) + \mathbf{b}_{\mathrm{base},2}, \end{aligned}

where Wbase1W_{\mathrm{base}}^{1}, Wbase2W_{\mathrm{base}}^{2} and biases bbase,1,bbase,2\mathbf{b}_{\mathrm{base},1}, \mathbf{b}_{\mathrm{base},2} are learnable parameters.

Delta adjustments for higher fidelities: For each subsequent fidelity level f>f0f > f_0, the model refines the prediction by adding a delta adjustment:

Δy(f)=fdelta,f(o(f))=Wf2⋅ϕ(Wf1⋅o(f)+bf,1)+bf,2.\Delta\mathbf{y}^{(f)} = f_{\mathrm{delta},f}\left(\mathbf{o}^{(f)}\right) = W_f^{2} \cdot\phi\left(W_f^{1} \cdot\mathbf{o}^{(f)} + \mathbf{b}_{f,1}\right) + \mathbf{b}_{f,2}.

The prediction at fidelity ff is then given by the sum of the base prediction and all applicable delta corrections:

y(f)=y(f0)+∑f′≤fΔy(f′).\mathbf{y}^{(f)} = \mathbf{y}^{(f_0)} + \sum_{f' \leq f} \Delta\mathbf{y}^{(f')}.

In this hierarchical structure, the shared attention mechanism with fidelity-specific queries enables each delta model to focus on different features of the shared representation, enhancing the model’s ability to capture fidelity-dependent nuances.

Bayesian Inference and Variational Training

Probabilistic model specification.— The Bayesian neural network (BNN) places a prior distribution over all its weights θ\theta and output noise standard deviations σ\sigma. For instance, we may use a half-normal prior for σ\sigma:

p(σ)=HalfNormal⁡(σ∣1.0),p(\sigma) = \operatorname{HalfNormal}(\sigma\mid1.0),

and standard normal priors for weights:

p(θ)=∏iN(θi∣0,1).p(\theta) = \prod_i \mathcal{N}(\theta_i \mid0,1).

Given input x\mathbf{x}, fidelity index ff, ZZ, and NN, the likelihood of observing target y\mathbf{y} is

p(y∣x,f,Z,N,θ,σ)=N(y;μ(x,f,Z,N;θ),diag⁡(σ2)),p(\mathbf{y} \mid\mathbf{x}, f, Z, N, \theta, \sigma) = \mathcal{N}\left(\mathbf{y}; \mu(\mathbf{x},f,Z,N;\theta), \operatorname{diag}(\sigma^2)\right),

where μ(⋅)\mu(\cdot) is the output of the hierarchical network.

Variational approximation and ELBO.— We approximate the intractable posterior p(θ,σ∣D)p(\theta,\sigma\mid\mathcal{D}) with a variational distribution q(θ,σ)q(\theta,\sigma), typically chosen as a mean-field Gaussian:

q(θ,σ)=∏iN(θi∣mi,si2)×q(σ),q(\theta,\sigma) = \prod_i \mathcal{N}(\theta_i \mid m_i,s_i^2) \times q(\sigma),

where mim_i and sis_i are variational parameters. For σ\sigma, we similarly choose a tractable form.

The Evidence Lower Bound (ELBO) over the dataset D\mathcal{D} is

L(q)=Eq(θ,σ)[∑(x,y,f,Z,N)∈Dlog⁡p(y∣x,f,Z,N,θ,σ)]−KL⁡(q(θ,σ)∥p(θ,σ)).\mathcal{L}(q) = \mathbb{E}_{q(\theta,\sigma)}\left[\sum_{(\mathbf{x},\mathbf{y},f,Z,N)\in\mathcal{D}} \log p(\mathbf{y} \mid\mathbf{x},f,Z,N,\theta,\sigma)\right] - \operatorname{KL}\left(q(\theta,\sigma)\mathbin{\|}p(\theta,\sigma)\right).

Loss components.— For a mini-batch B\mathcal{B} sampled from fidelity level ff, the negative ELBO loss function decomposes as:

LB=−∑i∈Blog⁡p(yi∣xi,f,Zi,Ni,θ,σ)+KL⁡(q(θ,σ)∥p(θ,σ)).\mathcal{L}_{\mathcal{B}} = -\sum_{i\in\mathcal{B}}\log p(\mathbf{y}_i \mid\mathbf{x}_i,f,Z_i,N_i,\theta,\sigma) + \operatorname{KL}\left(q(\theta,\sigma)\mathbin{\|}p(\theta,\sigma)\right).

Reconstruction Loss

Regularization

Optionally, a contrastive loss term Lcontrastive\mathcal{L}_{\mathrm{contrastive}} can be added to encourage similar embeddings for similar nuclei or fidelities:

Ltotal=LB+λLcontrastive,\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathcal{B}} + \lambda\mathcal{L}_{\mathrm{contrastive}},

with λ\lambda a hyperparameter weighting the contrastive component.

Training Protocol

The training procedure involves iteratively updating the variational parameters to maximize the ELBO:

  1. Initialization:

    • Set random seeds for reproducibility.

    • Initialize network weights, biases, and variational parameters.

    • Load multi-fidelity isotopic data, then preprocess and split into training, validation, and test sets.

    • Standardize input features x\mathbf{x} and target variables y\mathbf{y} using global scalers.

  1. Stochastic Variational Inference Loop:

    • Sample mini-batches Bf\mathcal{B}_f from each fidelity level ff.

  • Compute gradients of the negative ELBO LBf\mathcal{L}_{B_f} with respect to variational parameters.

  • Update parameters using an optimizer like ClippedAdam.

  1. Early Stopping and Checkpointing:

  • Track the validation loss and apply early stopping based on patience.

  • Save the model corresponding to the lowest validation loss.

The number of samples per each one of the fidelities and isotopes used in the training procedure is shown in Figure 7.

Samples per fidelity

Figure 7. Samples per fidelity: Number of LEC samples per fidelity used for training BANNANE for each one of the isotopes.

Inference and Uncertainty Quantification

To predict for a new input (x∗,f∗,Z∗,N∗)(\mathbf{x}^*, f^*, Z^*, N^*), we sample from the posterior predictive distribution following:

  1. Preprocessing: Scale x∗\mathbf{x}^* using the fitted StandardScaler.

  1. Monte Carlo Sampling: Draw SS samples {(θ(s),σ(s))}s=1S\{(\theta^{(s)},\sigma^{(s)})\}_{s=1}^{S} and compute predictions:

y∗(s)=μ(x∗,f∗,Z∗,N∗;θ(s)).\mathbf{y}^{*(s)} = \boldsymbol{\mu}(\mathbf{x}^*, f^*, Z^*, N^*; \theta^{(s)}).
  1. Estimating Statistics:

y^∗=1S∑s=1Sy∗(s),σ^∗=1S−1∑s=1S(y∗(s)−y^∗)2.\hat{\mathbf{y}}^* = \frac{1}{S}\sum_{s=1}^{S}\mathbf{y}^{*(s)}, \qquad\hat{\sigma}^* = \sqrt{\frac{1}{S-1}\sum_{s=1}^{S}\left(\mathbf{y}^{*(s)}-\hat{\mathbf{y}}^*\right)^2}.

Hyperparameters

To achieve the results presented in this work, we used a shared latent dimension of 80 with hidden layers of 64 units and a multi-head attention mechanism with 2 heads. The neutron number was encoded using a sinusoidal encoding, while the fidelity levels were embedded in an 8-dimensional space. The model included a dropout rate of 0.05 to enhance generalization. Training was performed using the ClippedAdam optimizer with an initial learning rate of 1×10−41 \times10^{-4}, and a patience of 200 for early stopping. The learning rate was reduced by a factor of 2 if validation loss did not improve within 25 evaluations. The number of training iterations was set to 30,000, with validation and test splits of 20% each. To account for the varying importance of different fidelities, we applied weighted losses, with higher fidelity levels receiving greater weight (emax⁡=4e_{\max}=4 to 10 assigned weights from 1.0 to 2.5).

With this configuration for the architecture, the total number of trainable parameters of the architecture is 2⋅1052 \cdot10^5.

Results

Complete Oxygen Chain

To assess the benefits of the global fitting performed by BANNANE, we compare the global fitting with a local fit performed using a state-of-the-art Boosted Decision Trees (BDT) [56] trained on a per-isotope basis including all fidelities, where the fidelity was included as an extra categorical feature. Table I presents a comparison of the global Root-mean-square error (RMSE) per isotope for both approaches.

NBANNANEBDT
EBE_B (MeV)RchR_{\mathrm{ch}} (fm)EBE_B (MeV)RchR_{\mathrm{ch}} (fm)
40.692(110)0.0318(8)2.521(89)0.0494(3)
50.415(72)0.0248(11)3.256(14)0.0498(2)
60.642(14)0.0131(3)4.541(90)0.0285(1)
70.544(12)0.0136(2)3.880(16)0.0240(1)
80.337(55)0.0205(7)5.535(15)0.0300(1)
90.317(45)0.0082(2)3.619(18)0.0239(1)
100.551(11)0.0116(2)5.294(11)0.0339(1)
120.455(66)0.0106(2)4.892(31)0.0246(1)
130.293(32)0.0083(1)8.109(30)0.0409(1)
140.277(32)0.0199(5)7.931(44)0.0450(2)
150.385(57)0.0101(1)6.318(24)0.0364(1)
160.538(11)0.0087(1)5.182(27)0.0340(1)

Table 1. Root-mean-square error (RMSE) of binding energy (EBE_B) and charge radius (RchR_{\mathrm{ch}}) for oxygen isotopes at emax⁡=10e_{\max}=10.

Leave One Out - Zero Shot extrapolation

As a tool to probe the extrapolation capabilities of BANNANE, we proceeded to remove completely each one of the isotopes from the training samples, and then we proceeded to use the model to evaluate the Mean Absolute Percentage Error (MAPE) for each one of them in a Zero-Shot approach, meaning that no examples for this fidelity were provided before. The distribution of residuals is shown in Figure 8. Remarkably, the model does not diverge significantly near the driplines, and the MAPEs do not increase considerably for higher emax⁡e_{\max} due to self-consistency from the hierarchical architecture.

Zero-Shot Leave-On-Out heatmaps

Figure 8. Zero-Shot Leave-On-Out. Distribution of the Zero-Shot extrapolation for unseen isotopes as each isotope was removed from the training Samples.

We find however that the model struggles with identifying the position of the nuclear shell closure if it isn’t included in the training, as seen by the significant error in the charge radii when removing one isotope near the edge of the shells. This means that it is important to add nuclei near the edges of the shell for the model to perform well. This is luckily not an issue for many-body methods as these nuclei are simpler to compute. Nuclear shells could be given as an input of the model in the future to remedy this.

Fidelity Extrapolation

Following the same approach as with the Leave One Out - Zero Shot extrapolation presented above, we repeated this experiment but now including only data with emax⁡=4e_{\max}=4 for each one of the isotopes left out and the rest of the fidelities for the rest of the chain. In Figure 9 we show the evaluated MAPE as extrapolated.

Low-Fidelity-Shot Leave-On-Out heatmaps

Figure 9. Low-Fidelity-Shot Leave-On-Out. Distribution of the emax⁡e_{\max} extrapolation for unseen isotopes as data from emax⁡>4e_{\max} > 4 was removed from the training Samples.

Interestingly, adding the low-fidelity samples considerably helps reduce the MAPE, while difficulties in predicting the dynamics of LECs around the shell closure remain for the RchR_{ch} regression.

Data Efficiency

Table 2 summarizes the data usage and resulting RMSE, underscoring the computational savings and the robust uncertainty quantification still retained by the Bayesian framework. This model refinement significantly cuts computational cost: we can cheaply train the base model on large amounts of emax⁡=4e_{\max}=4 data and only sample a sparse set of high-emax⁡e_{\max} which can be computed until the emulator residual converges.

emax⁡=8e_{\max}=8 data%MAPE BEB_E (%)MAPE RchR_{\mathrm{ch}} (%)
5%2.381.67
10%1.651.24
20%1.320.818
30%1.520.557
40%1.130.499
50%1.160.362
60%2.190.308
70%0.6500.326
80%0.7340.388
90%0.5710.267

Table 2. Inference (emax⁡=4,6e_{\max}=4,6) to high fidelity (emax⁡=8e_{\max}=8) with minimal extra data. Shown is the MAPE (%) in binding energies and charge radii for oxygen isotopes after introducing a given fraction of the emax⁡=8e_{\max}=8 dataset.

Ablation Study

In order to model the behavior of the model as more attention-exclusive parameters are added, we performed a sweep among the shared dimension layer of the LEC before passing them to the attention mask, while keeping all of the other parameters of the model fixed. For each one of the data points we performed 10 trainings with different random seeds, and for each one of them evaluated the corresponding test samples. In Figure 10, the robustness of the interpolation performance is evident. The attention mechanism allows the model to capture higher-order correlations between the LECs, allowing for globally better emulation of both the binding energy and the charge radii.

Ablation study of the size of the attention mechanism

Figure 10. Ablation study of the size of the attention mechanism as we systematically increased the size of the Multihead Attention layer while keeping the other parameters fixed.

General Sensitivity Analysis

Global sensitivity analysis is a statistical tool used to decompose how the total uncertainty of a model is separated in term of the uncertainty of its different parameters [49]. Following notation from [29], the total variance of the results, Var⁡[Y]\operatorname{Var}[Y], is decomposed as

Var⁡[Y]=∑i=1NLECsVi+∑jNLECs∑i<jNLECsVij+…(1)\operatorname{Var}[Y] = \sum_{i=1}^{N_{\mathrm{LECs}}} V_i + \sum_j^{N_{\mathrm{LECs}}} \sum_{i<j}^{N_{\mathrm{LECs}}} V_{ij} + \ldots \tag*{(1)}

with the partial variances Vi,Vij,…V_i,V_{ij},\ldots given by

Vi=Var⁡[Eα⃗∼(αi)[Y∣αi]](2)V_i = \operatorname{Var}\left[E_{\vec{\alpha}\sim(\alpha_i)}[Y|\alpha_i]\right] \tag*{(2)}
Vij=Var⁡[Eα⃗∼(αi,αj)[Y∣αi,αj]]−Vi−Vj.(3)V_{ij} = \operatorname{Var}\left[E_{\vec{\alpha}\sim(\alpha_i,\alpha_j)}[Y|\alpha_i,\alpha_j]\right] - V_i - V_j. \tag*{(3)}

α⃗∼(αi)\vec{\alpha}\sim(\alpha_i) is the set of LECs without αi\alpha_i and Eα⃗∼(αi)[Y∣αi]E_{\vec{\alpha}\sim(\alpha_i)}[Y|\alpha_i] is the conditional expectation of YY given αi\alpha_i, likewise for the higher order expression. We can then write the total sensitivity of the final results to each LEC αi\alpha_i as

ST(αi)=Si+Sij+Sijk+…(4)S_T(\alpha_i) = S_i + S_{ij} + S_{ijk} + \ldots \tag*{(4)}

where the partial sensitivities are given by

Si=ViVar⁡[Y],Sij=VijVar⁡[Y](5)S_i = \frac{V_i}{\operatorname{Var}[Y]}, \qquad S_{ij} = \frac{V_{ij}}{\operatorname{Var}[Y]} \tag*{(5)}

and likewise for higher orders.

Supplemental Material: Global Framework for Emulation of Nuclear Calculations

Antoine Belley; Jose M. Munoz; Ronald F. Garcia Ruiz

METHODS

Architecture Details

BANNANE is designed to emulate nuclear properties across isotopic chains while integrating multi-fidelity data. Its hierarchical Bayesian framework combines learnable embeddings, a multi-head self-attention mechanism with fidelity-specific queries, and fidelity-specific prediction pathways. This architecture ensures that the emulator can leverage correlations between nuclear observables across elements and isotopes, while also adapting to varying levels of computational fidelity. It must be added that we expect the model to further improve the performance once a full hyperparameter optimization is performed. Below we present a detailed overview of our implementation.

Input embeddings.— To incorporate discrete nuclear identifiers and fidelity information, BANNANE utilizes three primary embedding layers:

Element (ZZ) embedding: An embedding layer maps each proton number Z∈{0,1,…,Zmax⁡}Z \in\{0,1,\ldots,Z_{\max}\} to a continuous vector space:

eZ=WZ⋅1Z,\mathbf{e}_{Z} = \mathbf{W}_{Z} \cdot\mathbf{1}_{Z},

where WZ∈R(Zmax⁡+1)×dZ\mathbf{W}_{Z} \in\mathbb{R}^{(Z_{\max}+1)\times d_{Z}} is a learnable weight matrix, 1Z\mathbf{1}_{Z} is a one-hot encoding of ZZ, and dZd_{Z} is the embedding dimension. To increase the architecture’s flexibility, we left the option to include this embedding into the model, which is not considered for the benchmarks presented in this work but will be needed when other ZZ values are considered.

Neutron number (NN) positional encoding: A similar embedding can be applied to the neutron number NN: eN=WN⋅1N\mathbf{e}_{N} = \mathbf{W}_{N}\cdot\mathbf{1}_{N} with WN∈R(Nmax⁡+1)×dN\mathbf{W}_{N} \in\mathbb{R}^{(N_{\max}+1)\times d_{N}} and dNd_{N} the corresponding dimension. This learned positional encoding allows the network to capture trends as a function of neutron number, crucial for modeling isotopic evolution.

However, instead of a learnable embedding, BANNANE uses a fixed sinusoidal positional encoding inspired by Transformer architectures. This encoding defines a phase:

φ=⋅exp⁡(−2idNlog⁡(10000)),i=0,…,⌊dN−22⌋\varphi= \cdot\exp\left(-\frac{2i}{d_{N}}\log(10000)\right), \qquad i=0,\ldots,\left\lfloor\frac{d_{N}-2}{2}\right\rfloor

Which is then added as a positional encoding via:

PE⁡(n,2i)=sin⁡(nφ),PE⁡(n,2i+1)=cos⁡(nφ).\operatorname{PE}(n,2i)=\sin(n\varphi), \qquad\operatorname{PE}(n,2i+1)=\cos(n\varphi).

This guarantees that the neutron number encoding captures smooth trends across isotopic chains, aiding generalization.

The sinusoidal encoding is implemented as a configurable alternative to learned embeddings, selectable via hyperparameters. We used the sinusoidal encoded NN embeddings for all the results in this work.

Fidelity level embedding: Each fidelity level f∈Ff \in\mathcal{F} is similarly embedded:

ef=Wf⋅1f,\mathbf{e}_{f} = \mathbf{W}_{f}\cdot\mathbf{1}_{f},

where Wf∈R∣F∣×df\mathbf{W}_{f} \in\mathbb{R}^{|\mathcal{F}|\times d_{f}} maps a one-hot representation of fidelity level to an embedding space of dimension dfd_{f}.

The continuous input features x∈Rdx\mathbf{x}\in\mathbb{R}^{d_{x}}, which may include coupling constants and other parameters derived from effective field theories, are concatenated with these embeddings to form a combined input:

z=[x, eZ, eN, ef]∈Rdx+dZ+dN+df.\mathbf{z} = [\mathbf{x},\ \mathbf{e}_{Z},\ \mathbf{e}_{N},\ \mathbf{e}_{f}] \in\mathbb{R}^{d_{x}+d_{Z}+d_{N}+d_{f}}.

Shared latent representation.— The combined input z\mathbf{z} is fed into a shared latent network to extract a feature-rich representation:

h=ϕ(W2⋅ϕ(W1⋅z+b1)+b2),\mathbf{h} = \phi(\mathbf{W}_{2}\cdot\phi(\mathbf{W}_{1}\cdot\mathbf{z}+\mathbf{b}_{1})+\mathbf{b}_{2}),

where W1,W2\mathbf{W}_{1},\mathbf{W}_{2} are weight matrices, b1,b2\mathbf{b}_{1},\mathbf{b}_{2} are biases, and ϕ(⋅)\phi(\cdot) denotes a non-linear activation function (e.g., LeakyReLU). This produces a latent vector h∈Rdh\mathbf{h}\in\mathbb{R}^{d_{h}} shared across all fidelities.

Multi-head self-attention mechanism with fidelity-specific queries.— To capture complex interdependencies and fidelity-specific features, BANNANE employs a multi-head self-attention mechanism augmented with fidelity-specific query vectors [?]. This allows the model to extract different aspects of the shared representation tailored to each fidelity level. For our implementation, we fixed the number of attention heads to 2.

Fidelity-specific queries: For each fidelity level, including the base fidelity, a unique learnable query vector is introduced:

Q(f)=Wq(f)⋅h,\mathbf{Q}^{(f)}=\mathbf{W}_{q}^{(f)}\cdot\mathbf{h},

where Wq(f)\mathbf{W}_{q}^{(f)} is a learnable projection matrix specific to fidelity ff.

Attention computation: For a given fidelity ff, the query Q(f)\mathbf{Q}^{(f)} interacts with the shared latent representation h\mathbf{h} to compute attention weights and extract relevant features:

Attention⁡(f)(Q(f),K,V)=softmax⁡(Q(f)KTdk)V,\operatorname{Attention}^{(f)}(\mathbf{Q}^{(f)},\mathbf{K},\mathbf{V})=\operatorname{softmax}\left(\frac{\mathbf{Q}^{(f)}\mathbf{K}^{\mathsf{T}}}{\sqrt{d_{k}}}\right)\mathbf{V},

where K=Wkh\mathbf{K}=\mathbf{W}_{k}\mathbf{h} and V=Wvh\mathbf{V}=\mathbf{W}_{v}\mathbf{h} are shared key and value projections, respectively, and dkd_{k} is the dimension per head.

Concatenation and projection: The outputs from all attention heads are concatenated and passed through a linear layer to form the final attention-enhanced representation:

o(f)=Wo[head⁡1(f),…,head⁡H(f)],\mathbf{o}^{(f)} = W_o\left[\operatorname{head}^{(f)}_1,\ldots,\operatorname{head}^{(f)}_H\right],

with Wo∈R(H⋅dk)×doutW_o \in\mathbb{R}^{(H \cdot d_k) \times d_{\mathrm{out}}}.

Fidelity-specific regression pathways.— The attention-enhanced representation o(f)\mathbf{o}^{(f)} is used to predict target observables hierarchically, leveraging both shared and fidelity-specific information:

Base prediction: For the lowest fidelity level f0f_0, the model computes

y(f0)=fbase(o(f0))=Wbase2⋅ϕ(Wbase1⋅o(f0)+bbase,1)+bbase,2,\begin{aligned} \mathbf{y}^{(f_0)} = f_{\mathrm{base}}\left(\mathbf{o}^{(f_0)}\right) \\ &= W_{\mathrm{base}}^2 \cdot\phi\left(W_{\mathrm{base}}^1 \cdot\mathbf{o}^{(f_0)} + \mathbf{b}_{\mathrm{base},1}\right) + \mathbf{b}_{\mathrm{base},2}, \end{aligned}

where Wbase1W_{\mathrm{base}}^1, Wbase2W_{\mathrm{base}}^2 and biases bbase,1,bbase,2\mathbf{b}_{\mathrm{base},1}, \mathbf{b}_{\mathrm{base},2} are learnable parameters.

Delta adjustments for higher fidelities: For each subsequent fidelity level f>f0f > f_0, the model refines the prediction by adding a delta adjustment:

Δy(f)=fdelta,f(o(f))=Wf2⋅ϕ(Wf1⋅o(f)+bf,1)+bf,2.\Delta\mathbf{y}^{(f)} = f_{\mathrm{delta},f}\left(\mathbf{o}^{(f)}\right) = W_f^2 \cdot\phi\left(W_f^1 \cdot\mathbf{o}^{(f)} + \mathbf{b}_{f,1}\right) + \mathbf{b}_{f,2}.

The prediction at fidelity ff is then given by the sum of the base prediction and all applicable delta corrections:

y(f)=y(f0)+∑f′≤fΔy(f′).\mathbf{y}^{(f)} = \mathbf{y}^{(f_0)} + \sum_{f' \leq f} \Delta\mathbf{y}^{(f')}.

In this hierarchical structure, the shared attention mechanism with fidelity-specific queries enables each delta model to focus on different features of the shared representation, enhancing the model’s ability to capture fidelity-dependent nuances.

Bayesian Inference and Variational Training

Probabilistic model specification.— The Bayesian neural network (BNN) places a prior distribution over all its weights θ\theta and output noise standard deviations σ\sigma. For instance, we may use a half-normal prior for σ\sigma:

p(σ)=HalfNormal⁡(σ∣1.0),p(\sigma) = \operatorname{HalfNormal}(\sigma\mid1.0),

and standard normal priors for weights:

p(θ)=∏iN(θi∣0,1).p(\theta) = \prod_i \mathcal{N}(\theta_i \mid0,1).

Given input x\mathbf{x}, fidelity index ff, ZZ, and NN, the likelihood of observing target y\mathbf{y} is

p(y∣x,f,Z,N,θ,σ)=N(y;μ(x,f,Z,N;θ),diag⁡(σ2)),p(\mathbf{y} \mid\mathbf{x},f,Z,N,\theta,\sigma) = \mathcal{N}\left(\mathbf{y};\mu(\mathbf{x},f,Z,N;\theta),\operatorname{diag}(\sigma^2)\right),

where μ(⋅)\mu(\cdot) is the output of the hierarchical network.

Variational approximation and ELBO.— We approximate the intractable posterior p(θ,σ∣D)p(\theta,\sigma\mid\mathcal{D}) with a variational distribution q(θ,σ)q(\theta,\sigma), typically chosen as a mean-field Gaussian:

q(θ,σ)=∏iN(θi∣mi,si2)×q(σ),q(\theta,\sigma) = \prod_i \mathcal{N}(\theta_i \mid m_i,s_i^2) \times q(\sigma),

where mim_i and sis_i are variational parameters. For σ\sigma, we similarly choose a tractable form.

The Evidence Lower Bound (ELBO) over the dataset D\mathcal{D} is

L(q)=Eq(θ,σ)[∑(x,y,f,Z,N)∈Dlog⁡p(y∣x,f,Z,N,θ,σ)]−KL⁡(q(θ,σ)∥p(θ,σ)).\mathcal{L}(q) = \mathbb{E}_{q(\theta,\sigma)}\left[\sum_{(\mathbf{x},\mathbf{y},f,Z,N)\in\mathcal{D}} \log p(\mathbf{y}\mid\mathbf{x},f,Z,N,\theta,\sigma)\right] - \operatorname{KL}\left(q(\theta,\sigma)\mathbin\Vert p(\theta,\sigma)\right).

Loss components.— For a mini-batch B\mathcal{B} sampled from fidelity level ff, the negative ELBO loss function decomposes as:

LB=−∑i∈Blog⁡p(yi∣xi,f,Zi,Ni,θ,σ)+KL⁡(q(θ,σ)∥p(θ,σ)).\mathcal{L}_{\mathcal{B}} = -\sum_{i\in\mathcal{B}}\log p(\mathbf{y}_i\mid\mathbf{x}_i,f,Z_i,N_i,\theta,\sigma) + \operatorname{KL}\left(q(\theta,\sigma)\mathbin\Vert p(\theta,\sigma)\right).

Reconstruction Loss

Regularization

Optionally, a contrastive loss term Lcontrastive\mathcal{L}_{\mathrm{contrastive}} can be added to encourage similar embeddings for similar nuclei or fidelities:

Ltotal=LB+λLcontrastive,\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathcal{B}} + \lambda\mathcal{L}_{\mathrm{contrastive}},

with λ\lambda a hyperparameter weighting the contrastive component.

Training Protocol

The training procedure involves iteratively updating the variational parameters to maximize the ELBO:

  1. Initialization:

  • Set random seeds for reproducibility.

  • Initialize network weights, biases, and variational parameters.

  • Load multi-fidelity isotopic data, then preprocess and split into training, validation, and test sets.

  • Standardize input features x\mathbf{x} and target variables y\mathbf{y} using global scalers.

  1. Stochastic Variational Inference Loop:

  • Sample mini-batches Bf\mathcal{B}_f from each fidelity level ff.

  • Compute gradients of the negative ELBO LBf\mathcal{L}_{\mathcal{B}_f} with respect to variational parameters.

  • Update parameters using an optimizer like ClippedAdam.

  1. Early Stopping and Checkpointing:

Samples per fidelity heatmap

Figure 1 (PDF p. 15). Samples per fidelity: Number of LEC samples per fidelity used for training BANNANE for each one of the isotopes.

  • Track the validation loss and apply early stopping based on patience.

  • Save the model corresponding to the lowest validation loss.

The number of samples per each one of the fidelities and isotopes used in the training procedure is shown in Fig. 1.

Inference and Uncertainty Quantification

To predict for a new input (x∗,f∗,Z∗,N∗)(\mathbf{x}^{*}, f^{*}, Z^{*}, N^{*}), we sample from the posterior predictive distribution following:

  1. Preprocessing: Scale x∗\mathbf{x}^{*} using the fitted Standard-Scaler.

  1. Monte Carlo Sampling: Draw SS samples {(θ(s),σ(s))}s=1S\{(\boldsymbol{\theta}^{(s)}, \boldsymbol{\sigma}^{(s)})\}_{s=1}^{S} and compute predictions:

y∗(s)=μ(x∗,f∗,Z∗,N∗;θ(s)).\mathbf{y}^{*(s)} = \mu(\mathbf{x}^{*}, f^{*}, Z^{*}, N^{*}; \boldsymbol{\theta}^{(s)}).
  1. Estimating Statistics:

y^∗=1S∑s=1Sy∗(s),σ^∗=1S−1∑s=1S(y∗(s)−y^∗)2.\hat{\mathbf{y}}^{*} = \frac{1}{S}\sum_{s=1}^{S}\mathbf{y}^{*(s)}, \qquad\hat{\boldsymbol{\sigma}}^{*} = \sqrt{\frac{1}{S-1}\sum_{s=1}^{S}\left(\mathbf{y}^{*(s)}-\hat{\mathbf{y}}^{*}\right)^{2}}.

Hyperparameters

To achieve the results presented in this work, we used a shared latent dimension of 80 with hidden layers of 64 units and a multi-head attention mechanism with 2 heads. The neutron number was encoded using a sinusoidal encoding, while the fidelity levels were embedded in an 8-dimensional space.

The model included a dropout rate of 0.05 to enhance generalization. Training was performed using the ClippedAdam optimizer with an initial learning rate of 1×10−41 \times10^{-4}, and a patience of 200 for early stopping. The learning rate was reduced by a factor of 2 if validation loss did not improve within 25 evaluations. The number of training iterations was set to 30,000, with validation and test splits of 20% each. To account for the varying importance of different fidelities, we applied weighted losses, with higher fidelity levels receiving greater weight (emax⁡=4e_{\max}=4 to 10 assigned weights from 1.0 to 2.5).

With this configuration for the architecture, the total number of trainable parameters of the architecture is 2⋅1052 \cdot10^{5}.

Results

Complete Oxygen Chain

To assess the benefits of the global fitting performed by BANNANE, we compare the global fitting with a local fit performed using a state-of-the-art Boosted Decision Trees (BDT) [? ] trained on a per-isotope basis including all fidelities, where the fidelity was included as an extra categorical feature. Table I (PDF p. 15) presents a comparison of the global Root-mean-square error (RMSE) per isotope for both approaches.

NBANNANE EBE_B (MeV)BANNANE RchR_{ch} (fm)BDT EBE_B (MeV)BDT RchR_{ch} (fm)
40.692(110)0.0318(8)2.521(89)0.0494(3)
50.415(72)0.0248(11)3.256(14)0.0498(2)
60.642(14)0.0131(3)4.541(90)0.0285(1)
70.544(12)0.0136(2)3.880(16)0.0240(1)
80.337(55)0.0205(7)5.535(15)0.0300(1)
90.317(45)0.0082(2)3.619(18)0.0239(1)
100.551(11)0.0116(2)5.294(11)0.0339(1)
120.455(66)0.0106(2)4.892(31)0.0246(1)
130.293(32)0.0083(1)8.109(30)0.0409(1)
140.277(32)0.0199(5)7.934(44)0.0450(2)
150.385(57)0.0101(1)6.318(24)0.0364(1)
160.538(11)0.0087(1)5.182(27)0.0340(1)

Table I (PDF p. 15). Root-mean-square error (RMSE) of binding energy (EBE_B) and charge radius (RchR_{ch}) for oxygen isotopes at emax⁡=10e_{\max}=10.

Leave One Out - Zero Shot extrapolation

As a tool to probe the extrapolation capabilities of BANNANE, we proceeded to remove completely each one of the isotopes from the training samples, and then we proceeded to use the model to evaluate the Mean Absolute Percentage Error (MAPE) for each one of them in a Zero-Shot approach, meaning that no examples for this fidelity were provided before. The distribution of residuals is shown in Fig. 2. Remarkably, the model does not diverge significantly near the driplines, and the MAPEs do not increase considerably for higher emax⁡e_{\max} due to self-consistency from the hierarchical architecture.

Four heatmaps showing zero-shot leave-one-out MAPE distributions

Figure 2 (PDF p. 16). Zero-Shot Leave-On-Out. Distribution of the Zero-Shot extrapolation for unseen isotopes as each isotope was removed from the training Samples.

Four heatmaps showing low-fidelity-shot leave-one-out MAPE distributions

Figure 3 (PDF p. 16). Low-Fidelity-Shot Leave-On-Out. Distribution of the emax⁡e_{\max} extrapolation for unseen isotopes as data from emax⁡>4e_{\max} > 4 was removed from the training Samples.

We find however that the model struggles with identifying the position of the nuclear shell closure if it isn’t included in the training, as seen by the significant error in the charge radii when removing one isotope near the edge of the shells. This means that it is important to add nuclei near the edges of the shell for the model to perform well. This is luckily not an issue for many-body methods as these nuclei are simpler to compute. Nuclear shells could be given as an input of the model in the future to remedy this.

Fidelity Extrapolation

Following the same approach as with the Leave One Out - Zero Shot extrapolation presented above, we repeated this experiment but now including only data with emax⁡=4e_{\max} = 4 for each one of the isotopes left out and the rest of the fidelities for the rest of the chain. In Fig. 3 we show the evaluated MAPE as extrapolated.

Interestingly, adding the low-fidelity samples considerably helps reduce the MAPE, while difficulties in predicting the dynamics of LECs around the shell closure remain for the RchR_{ch} regression.

emax⁡=8e_{\max}=8 data %MAPE BEB_E (%)MAPE RchR_{ch} (%)
5%2.381.67
10%1.651.24
20%1.320.818
30%1.520.557
40%1.130.499
50%1.160.362
60%2.190.308
70%0.6500.326
80%0.7340.388
90%0.5710.267

Table 2 (PDF p. 16). Inference (emax⁡=4,6e_{\max}=4,6) to high fidelity (emax⁡=8e_{\max}=8) with minimal extra data. Shown is the MAPE (%) in binding energies and charge radii for oxygen isotopes after introducing a given fraction of the emax⁡=8e_{\max}=8 dataset.

Data Efficiency

Table 2 summarizes the data usage and resulting RMSE, underscoring the computational savings and the robust uncertainty quantification still retained by the Bayesian framework. This model refinement significantly cuts computational cost: we can cheaply train the base model on large amounts of emax⁡=4e_{\max}=4 data and only sample a sparse set of high-emax⁡e_{\max} which can be computed until the emulator residual converges.

Ablation Study

In order to model the behavior of the model as more attention-exclusive parameters are added, we performed a sweep among the shared dimension layer of the LEC before passing them to the attention mask, while keeping all of the other parameters of the model fixed. For each one of the data points we performed 10 trainings with different random seeds, and for each one of them evaluated the corresponding test samples. In Figure 4, the robustness of the interpolation performance is evident. The attention mechanism allows the model to capture higher-order correlations between the LECs, allowing for globally better emulation of both the binding energy and the charge radii.

Ablation study plots showing MAPE for binding energy and charge radii as a function of shared latent dimension

Figure 4 (PDF p. 17). Ablation study of the size of the attention mechanism as we systematically increased the size of the Multihead Attention layer while keeping the other parameters fixed.

General Sensitivity Analysis

Global sensitivity analysis is a statistical tool used to decompose how the total uncertainty of a model is separated in term of the uncertainty of its different parameters [? ]. Following notation from [? ], the total variance of the results, Var⁡[Y]\operatorname{Var}[Y], is decomposed as

Var⁡[Y]=∑i=1NLECsVi+∑iNLECs∑j<iNLECsVij+⋯(6)\operatorname{Var}[Y] = \sum_{i=1}^{N_{\mathrm{LECs}}} V_i + \sum_i^{N_{\mathrm{LECs}}} \sum_{j<i}^{N_{\mathrm{LECs}}} V_{ij} + \cdots \tag*{(6)}

with the partial variances ViV_i, VijV_{ij}, ... given by

Vi=Var⁡[Eα⃗∼(αi)[Y∣αi]](7)V_i = \operatorname{Var}\left[\mathbb{E}_{\vec{\alpha}\sim(\alpha_i)}[Y|\alpha_i]\right] \tag*{(7)}
Vij=Var⁡[Eα⃗∼(αi,αj)[Y∣αi,αj]]−Vi−Vj.(8)V_{ij} = \operatorname{Var}\left[\mathbb{E}_{\vec{\alpha}\sim(\alpha_i,\alpha_j)}[Y|\alpha_i,\alpha_j]\right] - V_i - V_j. \tag*{(8)}

α⃗∼(αi)\vec{\alpha}\sim(\alpha_i) is the set of LECs without αi\alpha_i and Eα⃗∼(αi)[Y∣αi]\mathbb{E}_{\vec{\alpha}\sim(\alpha_i)}[Y|\alpha_i] is the conditional expectation of YY given αi\alpha_i, likewise for the higher order expression. We can then write the total sensitivity of the final results to each LEC αi\alpha_i as

ST(αi)=Si+Sij+Sijk+⋯(9)S_T(\alpha_i) = S_i + S_{ij} + S_{ijk} + \cdots \tag*{(9)}

where the partial sensitivities are given by

Si=ViVar⁡[Y],Sij=VijVar⁡[Y](10)S_i = \frac{V_i}{\operatorname{Var}[Y]}, \qquad S_{ij} = \frac{V_{ij}}{\operatorname{Var}[Y]} \tag*{(10)}

and likewise for higher orders.

References

  1. [3]H. Hergert, S. K. Bogner, T. D. Morris, S. Binder, A. Calci, J. Langhammer, and R. Roth, Ab initio multireference in-medium similarity renormalization group calculations of even calcium and nickel isotopes, Phys. Rev. C 90, 041302 (2014).DOI
  2. [4]K. Hebeler, J. Holt, J. Menéndez, and A. Schwenk, Nuclear forces and their impact on neutron-rich nuclei and neutron-rich matter, Annual Review of Nuclear and Particle Science 65, 457 (2015).DOI
  3. [5]V. Lapoux, V. Somà, C. Barbieri, H. Hergert, J. D. Holt, and S. R. Stroberg, Radii and binding energies in oxygen isotopes: A challenge for nuclear forces, Phys. Rev. Lett. 117, 052501 (2016).DOI
  4. [6]R. F. Garcia Ruiz, M. L. Bissell, K. Blaum, A. Ekström, N. Frömmgen, G. Hagen, M. Hammen, K. Hebeler, J. D. Holt, G. R. Jansen, M. Kowalska, K. Kreim, W. Nazarewicz, R. Neugart, G. Neyens, W. Nörtershäuser, T. Papenbrock, J. Papuga, A. Schwenk, J. Simonis, K. Wendt, and D. T. Yordanov, Unexpectedly large charge radii of neutron-rich calcium isotopes, Nature Physics 12, 594–598 (2016).DOI
  5. [7]G. Hagen, A. Ekström, C. Forssén, G. R. Jansen, W. Nazarewicz, T. Papenbrock, K. A. Wendt, S. Bacca, N. Barnea, B. Carlsson, C. Drischler, K. Hebeler, M. Hjorth-Jensen, M. Morelli, G. Orlandini, A. Schwenk, and J. Simonis, Neutron and weak-charge distributions of the 48ca nucleus, Nature Physics 12, 186 (2016).DOI
  6. [8]J. Simonis, S. R. Stroberg, K. Hebeler, J. D. Holt, and A. Schwenk, Saturation with chiral interactions and consequences for finite nuclei, Phys. Rev. C 96, 014303 (2017).DOI
  7. [9]T. Duguet, V. Somà, S. Lecluse, C. Barbieri, and P. Navrátil, Ab initio calculation of the potential bubble nucleus 34Si, Phys. Rev. C 95, 034319 (2017).DOI
  8. [10]B.-N. Lu, N. Li, S. Elhatisari, D. Lee, E. Epelbaum, and U.-G. Meißner, Essential elements for nuclear binding, Physics Letters B 797, 134863 (2019).DOI
  9. [11]H. N. Liu, A. Obertelli, P. Doornenbal, C. A. Bertulani, G. Hagen, J. D. Holt, G. R. Jansen, T. D. Morris, A. Schwenk, R. Stroberg, N. Achouri, H. Baba, F. Browne, D. Calvet, F. Château, S. Chen, N. Chiga, A. Corsi, M. L. Cortés, A. Delbart, J.-M. Gheller, A. Giganon, A. Gillibert, C. Hilaire, T. Isobe, T. Kobayashi, Y. Kubota, V. Lapoux, T. Motobayashi, I. Murray, H. Otsu, V. Panin, N. Paul, W. Rodríguez, H. Sakurai, M. Sasano, D. Steppenbeck, L. Stuhl, Y. L. Sun, Y. Togano, T. Uesaka, K. Wimmer, K. Yoneda, O. Aktas, T. Aumann, L. X. Chung, F. Flavigny, S. Franchoo, I. Gasparić, R.-B. Gerst, J. Gibelin, K. I. Hahn, D. Kim, T. Koiwai, Y. Kondo, P. Koseoglou, J. Lee, C. Lehr, B. D. Linh, T. Lokotko, M. MacCormick, K. Moschner, T. Nakamura, S. Y. Park, D. Rossi, E. Sahin, D. Sohler, P.-A. Söderström, S. Takeuchi, H. Törnqvist, V. Vaquero, V. Wagner, S. Wang, V. Werner, X. Xu, H. Yamada, D. Yan, Z. Yang, M. Yasuda, and L. Zanetti, How robust is the n = 34 subshell closure? first spectroscopy of 52Ar, Phys. Rev. Lett. 122, 072502 (2019).
  10. [12]P. Gysbers, G. Hagen, J. D. Holt, G. R. Jansen, T. D. Morris, P. Navrátil, T. Papenbrock, S. Quaglioni, A. Schwenk, S. R. Stroberg, and K. A. Wendt, Discrepancy between experimental and theoretical β-decay rates resolved from first principles, Nature Physics 15, 428 (2019).DOI
  11. [13]V. Somà, P. Navrátil, F. Raimondi, C. Barbieri, and T. Duguet, Novel chiral hamiltonian and observables in light and medium-mass nuclei, Phys. Rev. C 101, 014318 (2020).DOI
  12. [14]A. Belley, T. Miyagi, S. R. Stroberg, and J. D. Holt, Constraining neutrinoless double-beta decay matrix elements from ab initio nuclear theory, (2022), arXiv:2210.05809 [nucl-th].DOI
  13. [15]A. Ekström, C. Forssén, G. Hagen, G. R. Jansen, T. Papenbrock, and Z. H. Sun, How chiral forces shape neutron-rich ne and mg nuclei (2023), arXiv:2305.06955 [nucl-th].arxiv.org/abs/2305.06955
  14. [16]P. Arthuis, K. Hebeler, and A. Schwenk, Neutron-rich nuclei and neutron skins from chiral low-resolution interactions, (2024), arXiv:2401.06675 [nucl-th].DOI
  15. [17]A. Belley, J. M. Yao, B. Bally, J. Pitcher, J. Engel, H. Hergert, J. D. Holt, T. Miyagi, T. R. Rodríguez, A. M. Romero, S. R. Stroberg, and X. Zhang, Ab initio uncertainty quantification of neutrinoless double-beta decay in 76Ge, Phys. Rev. Lett. 132, 182502 (2024).DOI
  16. [18]B. R. Barrett, P. Navrátil, and J. P. Vary, Ab initio no core shell model, Progress in Particle and Nuclear Physics 69, 131 (2013).
  17. [19]G. Hagen, T. Papenbrock, M. Hjorth-Jensen, and D. J. Dean, Coupled-cluster computations of atomic nuclei, Reports on Progress in Physics 77, 096302 (2014).DOI
  18. [20]S. R. Stroberg, A. Calci, H. Hergert, J. D. Holt, S. K. Bogner, R. Roth, and A. Schwenk, Nucleus-dependent valence-space approach to nuclear structure, Phys. Rev. Lett. 118, 032502 (2017).DOI
  19. [21]J. M. Yao, B. Bally, J. Engel, R. Wirth, T. R. Rodríguez, and H. Hergert, Ab initio treatment of collective correlations and the neutrinoless double beta decay of 48Ca, Phys. Rev. Lett. 124, 232501 (2020).
  20. [22]V. Somà, Self-consistent green’s function theory for atomic nuclei, Frontiers in Physics 8, 10.3389/fphy.2020.00340 (2020).DOI
  21. [23]H. Hergert, A guided tour of ab initio nuclear many-body theory, Frontiers in Physics 8, 10.3389/fphy.2020.00379 (2020).DOI
  22. [24]A. Tichai, R. Roth, and T. Duguet, Many-body perturbation theories for finite nuclei, Frontiers in Physics 8, 10.3389/fphy.2020.00164 (2020).
  23. [25]A. Tichai, S. Knecht, A. Kruppa, O. Legeza, C. Moca, A. Schwenk, M. Werner, and G. Zarand, Combining the in-medium similarity renormalization group with the density matrix renormalization group: Shell structure and information entropy, Physics Letters B 845, 138139 (2023).
  24. [26]T. D. Morris, J. Simonis, S. R. Stroberg, C. Stumpf, G. Hagen, J. D. Holt, G. R. Jansen, T. Papenbrock, R. Roth, and A. Schwenk, Structure of the lightest tin isotopes, Phys. Rev. Lett. 120, 125503 (2018).DOI
  25. [27]A. R. Vernon, R. F. Garcia Ruiz, T. Miyagi, C. L. Binnersley, J. Billowes, M. L. Bissell, J. Bonnard, T. E. Colomiers, J. Dobaczewski, G. J. Farooq-Smith, K. T. Flanagan, G. Georgiev, W. Gins, R. P. de Groote, R. Heinke, J. D. Holt, J. H. Hustings, Á. Koszorús, D. L. Leimbach, K. M. Lynch, G. Neyens, S. R. Stroberg, S. G. Wilkins, X. F. Yang, and D. T. Yordanov, Nuclear moments of indium isotopes reveal abrupt change at magic number 82, Nature 607, 260 (2022).DOI
  26. [28]J. Karthein, C. M. Ricketts, R. F. Garcia Ruiz, J. Billowes, C. L. Binnersley, T. E. Cocolios, J. Dobaczewski, G. J. Farooq-Smith, K. T. Flanagan, G. Georgiev, W. Gins, R. P. de Groote, F. P. Gustafsson, J. D. Holt, A. Kanellakopoulos, Á. Koszorús, D. Leimbach, K. M. Lynch, T. Miyagi, W. Nazarewicz, G. Neyens, P. G. Reinhard, B. K. Sahoo, A. R. Vernon, S. G. Wilkins, X. F. Yang, and D. T. Yordanov, Electromagnetic properties of indium isotopes illuminate the doubly magic character of 100Sn, Nature Physics 20, 1719 (2024), arXiv:2310.15093 [nuc-ex].
  27. [29]A. Ekström and G. Hagen, Global sensitivity analysis of bulk properties of an atomic nucleus, Phys. Rev. Lett. 123, 252501 (2019), arXiv:1910.02922 [nucl-th].DOI
  28. [30]A. Boehnlein, M. Diefenthaler, N. Sato, M. Schram, V. Ziegler, C. Fanelli, M. Horth-Jensen, T. Horn, M. P. Kuchera, D. Lee, W. Nazarewicz, P. Ostroumov, K. Orginos, A. Poon, X.-N. Wang, A. Scheinker, M. S. Smith, and L.-G. Pang, Colloquium: Machine learning in nuclear physics, Rev. Mod. Phys. 94, 031003 (2022).
  29. [31]A. Belley, Probing beyond standard model physics through ab initio calculations of exotic weak processes in atomic nuclei, Ph.D. thesis, University of British Columbia (2024).DOI
  30. [32]D. Frame, R. He, I. Ipsen, D. Lee, D. Lee, and E. Rrapaj, Eigenvector continuation with subspace learning, Phys. Rev. Lett. 121, 032501 (2018).DOI
  31. [33]C. Drischler, J. A. Melendez, R. J. Furnstahl, A. J. Garcia, and X. Zhang, Bugeye guide to projection-based emulators in nuclear physics, Frontiers in Physics 10, 10.3389/fphy.2022.1092931 (2023).arxiv.org/abs/2212.04912
  32. [34]T. Duguet, A. Ekström, R. J. Furnstahl, S. König, and D. Lee, Colloquium: Eigenvector continuation and projection-based emulators, Rev. Mod. Phys. 96, 031002 (2024).
  33. [35]W. G. Jiang, C. Forssén, T. Djärv, and G. Hagen, Emulating ab initio computations of infinite nucleonic matter, Phys. Rev. C 109, 064314 (2024).DOI
  34. [36]S. Yoshida, Imsrg-net: A machine-learning-based solver for the in-medium similarity renormalization group method, Phys. Rev. C 108, 044303 (2023).
  35. [37]P. Giuliani, K. Godbey, V. Kejzlar, and W. Nazarewicz, Model orthogonalization and bayesian forecast mixing via principal component analysis, Phys. Rev. Res. 6, 033266 (2024).arxiv.org/abs/2405.10839
  36. [38]D. Lay, E. Flynn, S. A. Giuliani, W. Nazarewicz, and L. Neufcourt, Neural network emulation of spontaneous fission, Phys. Rev. C 109, 044305 (2024).DOI
  37. [39]S. R. Stroberg, H. Hergert, S. K. Bogner, and J. D. Holt, Nonempirical interactions for the nuclear shell model: An update, Annual Review of Nuclear and Particle Science 69, 307 (2019).DOI
  38. [40]S. R. Stroberg, IMSRG++ (2018).
  39. [41]N. Shimizu, T. Mizusaki, Y. Utsuno, and Y. Tsunoda, Thick-restart block Lanczos method for large-scale shell-model calculations, Computer Physics Communications 244, 372 (2019).DOI
  40. [42]E. B. Bingham, J. P. Chen, M. Jankowiak, F. Obermeyer, N. Pradhan, T. Karaletsos, R. Singh, P. Szerlip, P. Horsfall, and N. D. Goodman, Pyro: Deep universal probabilistic programming, Journal of machine learning research 20, 1 (2019).arxiv.org/abs/1810.09538
  41. [43]M. D. Hoffman and M. J. Johnson, Elbo surgery: yet another how to carve up the variational evidence lower bound, in Workshop in Advances in Approximate Bayesian Inference, NIPS, Vol. 1 (2016).
  42. [44]See supplemental material (2025).
  43. [45]L. Van der Maaten and G. Hinton, Visualizing data using t-sne., Journal of machine learning research 9 (2008).
  44. [46]S. R. Stroberg, J. D. Holt, A. Schwenk, and J. Simonis, Ab initio limits of atomic nuclei, Phys. Rev. Lett. 126, 022501 (2021).DOI
  45. [47]M. Wang, W. J. Huang, F. G. Kondev, G. Audi, and S. Naimi, The ame 2020 atomic mass evaluation (ii). tables, graphs and references, Chinese Physics C 45, 030003 (2021).
  46. [48]V. Lapoux, V. Somà, C. Barbieri, H. Hergert, J. Holt, and S. Stroberg, Radii and binding energies in oxygen isotopes: A challenge for nuclear forces, Physical Review Letters 117, 10.1103/physrevlett.117.052501 (2016).DOI
  47. [49]I. M. Sobol, Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates, Mathematics and computers in simulation 55, 271 (2001).DOI
  48. [50]X. Yang, S. Wang, S. Wilkins, and R. G. Ruiz, Laser spectroscopy for the study of exotic nuclei, Progress in Particle and Nuclear Physics 129, 104005 (2023).DOI
  49. [51]A. Brown et al., Motivations for early high-profile frib experiments, (2025), arXiv:arXiv:2410.06144 [nucl-th].arxiv.org/abs/2410.06144
  50. [52]Munozarazism, paper_o_banname (2025), accessed: 2025-02-20.
  51. [53]C. Sanderson and R. Curtin, Armadillo: a template-based c++ library for linear algebra, Journal of Open Source Software 1, 26 (2016).DOI
  52. [54]C. Sanderson and R. Curtin, A user-friendly hybrid sparse matrix class in c++, in Lecture Notes in Computer Science (LNCS), Vol. 10931 (2018) pp. 422–430.DOI
  53. [55]A. Vaswani, Attention is all you need, Advances in Neural Information Processing Systems (2017).arxiv.org/abs/1706.03762
  54. [56]L. Prokhorenkova, G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin, Catboost: unbiased boosting with categorical features, Advances in neural information processing systems 31 (2018).arxiv.org/abs/1706.09516

Paper details

vlnc:W_ycm79qup
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents