betaFeedback

WorksPhysics and Astronomy

Neural Network predictions of inclusive electron-nucleus cross sections

Abstract

We investigate whether a neural network approach can reproduce and predict the electron-nucleus cross sections in the kinematical domain of present and future accelerator-based neutrino oscillation experiments. For this purpose, we consider the large amount of data available to the community via the web-page ``Quasielastic Electron Nucleus scattering archive'', and use a residual, fully connected feedforward neural network. We illustrate the training performances of the neural network by comparing its results with experimental data for the electron double-differential cross section on carbon. The agreement between predictions and data is remarkable from quasielastic to deep-inelastic scattering. To test the predicting power of the neural network we consider the numerous kinematical conditions for which experimental cross sections on calcium are available. Furthermore, we show the predictions of the electron scattering cross sections on oxygen, argon, and titanium: nuclei of particular interest in the context of present and future accelerator-based neutrino oscillation program. The agreement between these predictions and the data is comparable to the one of other theoretical models commonly used to calculate electron and neutrino cross sections, such as SuSAv2 and GiBUU. Results obtained with GENIE, a Monte Carlo event generator, are also discussed for comparison. The good performances obtained with our neural network suggest that neural networks could be exploited for theoretical and experimental investigations of electron- and neutrino-nucleus scattering.

Introduction

For over a hundred years, the electron-scattering experiments have represented one of the most powerful and fruitful approaches for investigating the structure of physical systems. The Franck–Hertz experiment [1], which showed that the energy absorption by atoms is quantized by studying the flux of electrons through a vapor of mercury atoms can be considered as the progenitor of an important tradition of electron-scattering experiments allowing us to move the frontiers of our knowledge of the structure of matter: from atoms, to nuclei [2], to quarks, precisely discovered thanks to experiments [3, 4] of this kind.

The possibility of using other lepton beams also appeared in parallel with the development of increasingly energetic and intense electron beams. In particular, starting from Pontecorvo’s [5] idea and Schwartz’s [6] project, experiments with neutrino beams have been performed, the first having led to the discovery of muon neutrino [7].

Nowadays accelerator-based neutrino experiments, such as T2K [8] and NOvA [9], are performed for a precise determination of neutrino oscillation parameters. The next generation experiments DUNE [10] and Hyper-Kamiokande [11] will play a central role in the neutrino oscillation program.

To ensure the success of these experiments, a reduction in systematic errors to the level of a few percent is needed. Today, one of the greatest sources of systematic errors are the neutrino–nucleus cross sections. In these experiments nuclear targets (such as C, O, and Ar) are involved, and in the energy region of hundreds of MeV to a few GeV, these cross sections are known to a precision not exceeding 20 % [12, 13]. Knowledge of these cross sections is crucial to determine the neutrino energy that enters the expression of the neutrino oscillation probability. Since neutrino beams are not monochromatic, in contrast with electron beams, the initial neutrino energy is reconstructed from the final states of the neutrino-nucleus reaction.

The current program of electron-scattering experiments continues in connection with nuclear, hadronic and particle physics and represents a powerful tool in connection with neutrino physics. Indeed, electrons and neutrinos both being leptons, their interaction with atomic nuclei is similar: it happens via vector current in the case of electrons and via vector and vector-axial currents in the case of neutrinos. Hence the vector part of the cross section and the final state interaction of the scattered particles are identical for electron and neutrino scattering. Although the axial part of the neutrino cross section complicates the comparisons, the electron-scattering experiments, which have the great advantage of employing monoenergetic electron beams, are useful in the investigation of neutrino-nucleus scattering from several points of view. First of all, the large amount of cross section data already obtained by many different electron-scattering experiments can be used to validate microscopic theoretical models employed to predict neutrino-nucleus cross sections; a non exhaustive list of such studies includes the works of Refs. [16, 17] [14, 15]. These world data have been recently considered in Refs. [18, 19, 20] to benchmark the neutrino Monte Carlo event generators, such as GENIE [21], commonly used in the neutrino community, as well as to check the neutrino energy reconstruction via calorimetric and quasielastic kinematics-based methods [22]. The available data also includes the recently published electron cross sections on argon and titanium [23, 24, 25, 26]. These cross sections were measured in the experiment proposed in Ref. [27] to determine the spectral function of 40Ar{}^{40}\mathrm{Ar}, the nucleus employed in DUNE, and in the detectors of the Fermilab Short-baseline neutrino program [28].

The aim of this work is to investigate the use of neural networks to reproduce and predict the electron-nucleus cross sections in the kinematical domain of the present and future accelerator-based neutrino oscillation experiments.

In the last decade, applications using neural networks have become ubiquitous and are found in many tasks beyond fundamental research. In the high energy physics community, deep learning provides a faster alternative to both standard data analysis techniques and Monte Carlo approaches to simulate the detector outputs [29, 30, 31, 32, 33]. Deep learning for event reconstruction in accelerator-based neutrino experiments is used for example in Refs. [34, 35, 36, 37, 38, 39, 40]. Recent studies [41, 42] applied generative adversarial networks to construct an AI-based Monte Carlo event generator for deep-inelastic electron-proton scattering, free of theoretical assumptions about the underlying particle dynamics. In the context of electron and neutrino scattering, theoretical studies have used neural networks to obtain information on the two main ingredients of the lepton-nucleus cross sections: nuclear responses, and nucleon form factors. More precisely, in Ref.[43] a physics-informed artificial neural network is employed to reconstruct the electromagnetic response functions. In Ref. [44] a Bayesian approach for feed-forward neural networks has been applied to extract from the neutrino-deuteron scattering data the nucleon axial form factor, a quantity which has been widely debated in the last ten years, following the MiniBooNE measurement of the quasielastic-like neutrino cross section on carbon [45]. The possibility of using machine learning algorithms in reconstructing neutrino energy has been explored in Ref. [46]. For a recent review on the current trends and perspectives of artificial intelligence in nuclear physics we refer to [47].

4He{}^{4}\mathrm{He}6Li{}^{6}\mathrm{Li}9Be{}^{9}\mathrm{Be}12C{}^{12}\mathrm{C}16O{}^{16}\mathrm{O}24Mg{}^{24}\mathrm{Mg}27Al{}^{27}\mathrm{Al}40Ar{}^{40}\mathrm{Ar}40Ca{}^{40}\mathrm{Ca}48Ca{}^{48}\mathrm{Ca}48Ti{}^{48}\mathrm{Ti}56Fe{}^{56}\mathrm{Fe}59Ni{}^{59}\mathrm{Ni}
21.51.33.124.01.00.36.31.410.59.61.419.30.3

Table 1. Percent contribution of the different nuclear targets electron scattering data [50] considered in our study.

Data

A large amount of data is needed to predict electron-nucleus cross sections via a neural network approach. These data, accumulated by many different electron scattering experiments which began in the mid-1970s have been assembled by the authors of Refs. [48, 49] and made available in the “Quasielastic Electron Nucleus scattering archive” [50].

This archive contains about 600 different combinations of targets, energies and angles consisting of some 20000 data points. These data cover the energy region of interest for the present and future long-baseline neutrino oscillation experiments, from giant resonance excitations up to the deep-inelastic scattering, with a predominant contribution of quasielastic and Δ\Delta resonance excitations.

The nuclei included in the archive [50] vary from hydrogen to uranium. Driven by the idea of having a relatively homogeneous data-set, containing the nuclei employed in neutrino detectors, we decided to discard the very light nuclei and the heavy nuclei characterized by a large neutron-proton asymmetry. The final subset of nuclei considered varies between 4He{}^{4}\mathrm{He} and 59Ni{}^{59}\mathrm{Ni}. They are specified in Table I together with their relative contribution to our dataset.

In Figure 1 we plot some examples of inclusive double-differential cross sections d2σdωdΩ\frac{d^{2}\sigma}{d\omega d\Omega} for different values of incoming electron energy EE and electron scattering angles θ\theta1 as a function of the energy transferred to the nucleus ω\omega (also called energy loss, being ω=E−E′\omega= E - E' the difference between the incoming and outgoing electron energy E′E'. For this illustration we have selected 12C{}^{12}\mathrm{C}, a nucleus for which much data exists. Two characteristics that would complicate the neural network prediction task can already be seen from these few examples which refer to different kinematical conditions. Firstly, the cross section may span many orders of magnitude: 5 orders in the four panels of Figure 1, which refers to the same nucleus, and 12 orders of magnitude (from 4×10−64 \times10^{-6} to 10610^{6}) when considering the whole dataset. Secondly, the shapes of the cross section can vary greatly, reflecting the different reaction mechanisms that can be induced by different kinematics. Starting from the top left panel of

Image not extracted

Figure 1. Inclusive (e,e′)(e,e') 12C{}^{12}\mathrm{C} double-differential cross section data as a function of the transferred energy in four different kinematical conditions written above each panel. The vertical lines correspond to the kinematics of quasielastic and Δ\Delta excitations given by Eq. (2) and Eq. (4), respectively.

Figure 1, the first recognizable excitations are the nuclear giant resonances which correspond to the sharp peaks in the cross sections for ω<20\omega< 20 MeV. Moreover, the quasielastic bump appears in the three first panels. It corresponds to one nucleon knockout and is peaked around

ωQE=Q22MN=q2−ω22MN=q2+MN2−MN,(1)\omega_{QE} = \frac{Q^2}{2M_N} = \frac{\mathbf{q}^2-\omega^2}{2M_N} = \sqrt{\mathbf{q}^2+M_N^2}-M_N, \tag*{(1)}

where MNM_N is the nucleon mass and q\mathbf{q} the momentum transfer to the nucleus, given by the difference between the incoming and scattered electron momentum, q=k−k′\mathbf{q} = \mathbf{k} - \mathbf{k}'. In terms of electron kinematics variable, when electron mass is neglected, Eq. (1) can be written as:

ωQE=E2(1−cos⁡θ)MN+E(1−cos⁡θ).(2)\omega_{QE} = \frac{E^2(1-\cos\theta)}{M_N+E(1-\cos\theta)}. \tag*{(2)}

A vertical line corresponding to this value is plotted in each panel of Figure 1. The shift of the position of the real quasielastic peak with respect to the value of Eq. (1) and Eq. (2) is due to the absence of nucleon binding and nuclear collective effects in these formulae. The broadening of the quasielastic bump is due to nucleon Fermi motion. The second bump at a higher ω\omega corresponds to the Δ\Delta resonance excitation. In the case of scattering with a free nucleon at rest, it would peak at:

ωΔ=q2+MΔ2−MN=Q22MN+ΔM,(3)\omega_{\Delta} = \sqrt{\mathbf{q}^2+M_{\Delta}^2}-M_N = \frac{Q^2}{2M_N}+\Delta M, \tag*{(3)}

with ΔM=(MΔ2−MN2)/2MN=338\Delta M = (M_{\Delta}^2-M_N^2)/2M_N = 338 MeV. Equivalently, Eq. (3) can be written as:

ωΔ=MNΔM+E2(1−cos⁡θ)MN+E(1−cos⁡θ),(4)\omega_{\Delta} = \frac{M_N\Delta M+E^2(1-\cos\theta)}{M_N+E(1-\cos\theta)}, \tag*{(4)}

also shown in Figure 1. Beyond the Δ\Delta excitations, higher energy nucleon resonances contribute to the cross section up to the onset of the deep inelastic scattering. On the other hand, in the ω\omega region between the quasielastic and Δ\Delta peaks, the so called “dip” region, a large part of the cross section is due to multinucleon excitations, arising from nucleon-nucleon correlations and meson exchange processes. These complicated many-body mechanisms have attracted a lot of attention in the neutrino community in recent years, starting from the suggestion [51] of their inclusion as natural explanation of the MiniBooNE cross sections [45]. These mechanisms play a crucial role in the neutrino energy reconstruction problem [52, 53, 54, 55, 56].

In this work, we remove the small number of data points corresponding to the low-energy giant resonances. Their presence worsens the neural network predictions for essentially two reasons: they are characterized by the largest (and spiked) cross sections, and they represent a small proportion of the whole dataset, which focuses on quasielastic excitations and beyond. We postpone the inclusion of these excitations to future work.

Several microscopic quantum-mechanical and phenomenological approaches exist to describe the different reaction mechanisms that contribute to the inclusive electron-nucleus cross section. While some approaches allow a description of all the nuclear excitations mentioned earlier, others focus on specific excitations only. For a review of these models we refer the reader to Ref. [49] or to Ref. [57], where different approaches are discussed in connection with both electron and neutrino scattering. In the following we analyze to what extent neural networks can reproduce the electron-nucleus cross sections.

Neural Networks

In regression tasks, machine learning models are designed to learn mappings between an input domain and a predefined set of outputs for the target variables. The dataset considered in this work is composed of different features: the atomic number ZZ, the nucleus mass number AA, the electron beam energy EE, the electron scattering angle θ\theta, the energy loss ω\omega and the inclusive double-differential cross section d2σdωdΩ\frac{d^{2}\sigma}{d\omega d\Omega}. This last quantity is the value to be predicted by the supervised learning algorithm we present. Beyond the five data features (ZZ, AA, EE, θ\theta, and ω\omega), we add four additional complementary variables, obtained as combinations of the five original ones. These added variables are: cos⁡θ\cos\theta, ωQE\omega_{QE} defined in Eq. (2), ωΔ\omega_{\Delta} defined in Eq. (4), and Q2=q2−ω2Q^{2} = \mathbf{q}^{2} - \omega^{2}, here defined as:

Q2=2E(E−ω)(1−cos⁡θ),(5)Q^{2} = 2E(E-\omega)(1-\cos\theta), \tag*{(5)}

a relation that is only valid when the electron mass is set to zero.

The addition of handcrafted features, such as the four-momentum transfer in the reaction and the approximate position of the quasielastic and Δ\Delta peaks (despite a possible offset between real and approximated values due to the lack of removal energy parameters in Eqs. (2) and (4)), is found to be useful to drive the network.

A larger neural network is expected to perform equally well without adding handcrafted features but our choice to include these additional variables makes the optimization process easier and allows us to use a smaller network, that is faster to train and less prone to overfitting. However, exactly quantifying how these four variables actually improve the prediction is not possible since there is no guarantee that the fine tuning is equally well performed when we add the four additional variables and when we do not add them.

The model is built upon deep neural networks (DNNs) [58]. DNNs are designed to learn hierarchical and increasingly abstract representations of the data. A DNN is a composition of LL parametric functions named layers. The output of each layer, flf_l, l∈{0,…,L−1}l \in\{0,\ldots,L-1\}, is understood as a representation of the input samples. More specifically, the layer is composed of neurons, which are the building blocks of a layer. Hence, the layer flf_l takes the output of the previous layer fl−1f_{l-1} and applies a non-linear transformation to compute its output. These transformations use the model parameters, WlW_l for each layer, commonly called weights and conveniently represented by a rectangular matrix. They relate the neurons of a layer to the previous layer and contain the information extracted by the model from the training data. Thus, given an input xx, a neural network ff performs the following computation to infer its output:

f(W,x)=fL−1(WL−1,fL−2(WL−2,…,f0(W0,x))…).(6)f(W,x) = f_{L-1}(W_{L-1},f_{L-2}(W_{L-2},\ldots,f_0(W_0,x))\ldots). \tag*{(6)}

In Eq. (6), the state of layer ll can be denoted xlx^l and is called a representation. In a standard feedforward network, xl−1x^{l-1} is employed to compute xlx^l as follows:

xl=fl−1(Wl−1,xl−1)=g(Wl−1∗xl−1+bl−1),(7)x^l = f_{l-1}(W_{l-1},x^{l-1}) = g(W_{l-1} * x^{l-1} + b_{l-1}), \tag*{(7)}

where Wl−1∗xW_{l-1} * x is a matrix product, xx can be seen as a column vector of nl−1n_{l-1} components, then Wl−1W_{l-1} is a matrix of nl−1n_{l-1} columns and nln_l lines, bl−1b_{l-1} is a column vector of nln_l lines. gg is an element-wise nonlinear function. We used the standard Rectified Linear Unit, ReLU function defined as ReLU⁡(z)=max⁡{0,z}\operatorname{ReLU}(z) = \max\{0,z\}. Other choices might work equally well or even better, we make no claim of optimality in this paper. The result of Eq. (7) is the input of the next layer of the neural network, layer ll.

During the initial exploratory phase of this work we used fully connected neural networks as described by Eq. (7). We found that the richness of the data was such that deep neural networks with less than 35 layers were unable to fit the training set. We therefore used deeper networks that then became difficult to train for the reasons explained in Ref. [59]. The solution proposed by the authors of Ref. [59] is to use what they named a residual network. The architecture of our neural network is a fully connected residual network [59] which means that we perform an addition that shortcuts the architecture to help the optimization. procedure converge. The skipped connection is performed as follows:

xl=fl−1(Wl−1,xl−1,xl−5)=g(Wl−1∗g(xl−1+xl−5)+bl−1).(8)x^{l} = f_{l-1}(W_{l-1}, x^{l-1}, x^{l-5}) = g(W_{l-1} * g(x^{l-1} + x^{l-5}) + b_{l-1}). \tag*{(8)}

The skipped connections or residual blocks are used in deep learning to help training deep neural networks [58, 59]. Our neural network, implemented with TensorFlow [60], is composed of 10 residual blocks of 5 fully connected layers with increasing then decreasing sizes ranging from input size (9 input data) up to size 50 then down to the size 5 before connecting to the output of size one. Our neural network architecture is represented in Figure 2, the resulting model has 50 connected layers. To help information flow in the backpropagation stage we have chosen a layer width that varies progressively. Abrupt changes in widths made the optimization process harder and local minima were difficult to escape from. The constraint we have is that we start with 9 inputs and have an output of dimension 1.

Image not extracted

Figure 2. Illustration of the chosen neural network, composed of ten fully connected residual blocks. Details of one of these residual blocks are shown on the right of the figure. The activation function used is a Rectified linear unit (ReLU).

An essential element of neural networks training is the choice of the cost function. For the neural network to be able to accurately predict the cross section structure over its wide range and different shapes, we chose to use the relative absolute error given by:

C(y^)=1n∑i=1n∣yi−y^i∣yi.(9)C(\hat{y}) = \frac{1}{n} \sum_{i=1}^{n} \frac{\lvert y_i - \hat{y}_i \rvert}{y_i}. \tag*{(9)}

In Eq. (9), y^i\hat{y}_i is a model prediction and yy is the cross section divided by AA, y=d2σdωdΩ1Ay = \frac{\mathrm{d}^{2}\sigma}{\mathrm{d}\omega\mathrm{d}\Omega}\frac{1}{A} obtained from the dataset. In our case,

y^i=f(W,xi),(10)\hat{y}_i = f(W, x_i), \tag*{(10)}

where the input data is the nine dimensional vector of normalized components : xi=(Z,A,E,θ,ω,cos⁡θ,ωQE,ωΔ,Q2)x_i = (Z, A, E, \theta, \omega, \cos\theta, \omega_{QE}, \omega_{\Delta}, Q^{2}) and (xi,yi)(x_i, y_i) is an input-output pair. Another choice for the loss function could be the χ2\chi^{2} that would allow to take into account the experimental errors. For this first exploratory study, driven by a major simplicity and rapidity of calculations and by the fact that in the present data set the experimental errors are in general small, we prefer to chose the relative absolute error as cost function. We leave the use of χ2\chi^{2} to future works.

During the training phase, the model is presented with a large number of input-output pairs. The weights, WW, are initialized randomly with a normal distribution [58] and an initial prediction of the output is computed using Eq. (6). During the forward pass, the model’s prediction error is estimated by computing the value of a cost function that quantifies the discrepancy between the current prediction y^\hat{y} given by Eq. (9) and the true target yy. This cost function is computed over a batch of data and differentiated with respect to the model weights, WW, in the backward pass. Model weights are then updated with the computed differential to improve the predictions of the neural network. The general idea is to update WW in the following way:

W=W−ϵ∂C∂W.(11)W = W - \epsilon\frac{\partial C}{\partial W}. \tag*{(11)}

The variable ϵ\epsilon is called the learning rate and the value we chose was standard, ϵ=0.001\epsilon= 0.001. In practice, we use Adam optimizer [61] which provides an empirical improvement over Eq. (11) with a slowly decreasing learning rate, reminiscent of annealing, to help the optimization process. The values of the neural network weights that minimize the cost function are obtained during the training phase by iteratively taking forward and backward passes on the training data set. Once a predetermined number of epochs (a forward and a backward pass) is computed, the training phase is finished and the model is deployed on new samples, unseen during the training, to make predictions. Training takes about 30 minutes for 5000 epochs on a 12GB NVIDIA Tesla K80 GPU. We include standard L2L_2 regularization in order to reduce the variance of the model and hopefully improve generalization: the predictions on data not used during the training phase. Weights are stored during training and for inference we roll back to the epoch that yielded the lowest validation error during the whole training. In the inference phase, the output of the model is computed for the test data set. One hopes that the model will generalize well on the test data and infer target values close to the actual values thanks to the patterns it has learned from its training data.

Results

The training set is composed of the available samples excluding one or more chosen isotopes (Z′Z^\prime, A′A^\prime), or examples corresponding to some particular kinematical conditions. These excluded data will be examined in the testing stage. In other words, we used the trained neural network to make predictions for the isotopes (Z′Z^\prime, A′A^\prime) or for some specific kinematical conditions. This allows us to assess the ability of the trained neural network to infer the desired target for cases unseen during the training stage.

Training results

To further illustrate our procedure let us consider some specific cases. The whole dataset, representing our starting point, is the subset of the electron scattering data [50] corresponding to nuclei from 4He{}^{4}\mathrm{He} to 59Ni{}^{59}\mathrm{Ni}. Their relative contribution is illustrated in Table I. As a first experiment, we remove from the entire dataset only the isotope (Z=20Z = 20, A=40A = 40), i.e. the 40Ca{}^{40}\mathrm{Ca}, our test nucleus, for which we intend to predict the cross section. The rest of the dataset is randomly split: 90% for the training set and the remaining 10% for the validation set that allows us to monitor overfitting.

Stochastic gradient descent on minibatches is used for parameters optimization [62]. The minibatches of 512 instances of the training set are constituted randomly. This promotes regularization and speeds up the training. As a standard practice, at the end of the training process over 4000 epochs, we plot the evolution of the cost function, displayed in Figure 3, to check that the training has converged as expected and to monitor overfitting. In order to reduce the variance of the model and to control overfitting we used L2L_{2} regularization on the first and largest layers. It is standard practice to regularize large layers, that are more prone to overfitting, more than small layers. To limit overfitting we use for our prediction the parameters that gave the best validation value.

Image not extracted

Figure 3. Evolution of the cost function during the 4000 epochs training process for a successful training run. This illustrates a slight overfitting, controlled by L2L_{2} regularization.

In order to illustrate the overall performance of the neural network on the training set, we consider its predictions for 12C{}^{12}\mathrm{C}, a nucleus currently employed in the T2K and NOvA neutrino detectors. For this nucleus, much electron-scattering cross section data are available, for different kinematical conditions, inducing different nuclear excitations, from nuclear giant resonances up to deep inelastic scattering, as discussed in § 2.

The neural network double-differential cross section results for 12C{}^{12}\mathrm{C} are illustrated in Figure 4 and Figure 5 for a selected set of the values of incoming electron energy EE and electron scattering angle θ\theta present in the database. The overall agreement with data is remarkable for all the incoming energies and electron scattering angle2.

Image not extracted

Figure 4. Comparisons between the experimental inclusive (e,e′)(e,e') double differential cross section on 12C^{12}\mathrm{C}, in black squares with experimental error bars, and the neural network’s predictions, red circles, for the different values of incoming electron energy EE and scattering angle θ\theta (indicated above each panel), as a function of the transferred energy. To obtain these predictions, the 12C^{12}\mathrm{C} data occurrences were used in the training set as well as all the other nuclei of the dataset except 40Ca^{40}\mathrm{Ca}.

Image not extracted

Figure 5. The same as Figure 4 but for other kinematical conditions.

These good results suggest that neural networks can be considered as a tool for representing data that has this kind of behavior (changing shape depending on the type of induced excitation). In this context, we recall that for 12C{}^{12}\mathrm{C} the values of d2σdωdΩ\frac{d^{2}\sigma}{d\omega d\Omega} vary between a minimum of 7.7×10−5 nbGeV sr7.7 \times10^{-5}\,\frac{\mathrm{nb}}{\mathrm{GeV\,sr}} and a maximum of 9.6×105 nbGeV sr9.6 \times10^{5}\,\frac{\mathrm{nb}}{\mathrm{GeV\,sr}}. Despite this variation of 10 orders of magnitude, the choice of the mean relative absolute error of Eq. (9) as cost function allows for efficient training of the neural network.

Testing results

Following the results obtained for the training set, in this section we investigate the predictive power of the neural networks for the test set. First, we consider 40Ca{}^{40}\mathrm{Ca} as test nucleus. This choice is motivated by the existence of many available cross section data for this nucleus. The neural network predictions for this nucleus, which, we reiterate, was not included in the training set, are shown in Figure 6 and Figure 7. These figures show the result of an average over N=100N = 100 different neural networks trained on the same dataset. After initializing the same neural network architecture with NN different realizations of a random variable, we train the neural networks. For each training run we keep the weights that performed best on the validation set. These different neural networks give different predictions on the test set. For each scattering angle θ\theta, each energy EE, and each transferred energy ω\omega, we obtain NN independent values for y^i\hat{y}_i, the prediction for d2σdωdΩ\frac{d^{2}\sigma}{d\omega d\Omega}. We can then estimate the average based on these NN different predictions. Assuming a Gaussian distribution of these predictions, average and 95% confidence intervals for the true average μ\mu are estimated as follows:

Image not extracted

Figure 6. Comparisons between the experimental inclusive (e,e′)(e,e^\prime) double-differential cross section on 40Ca{}^{40}\mathrm{Ca}, in black squares with experimental error bars, and the neural network’s predictions, red circles, for the different values of incoming electron energy EE and scattering angle θ\theta written above each panel, as a function of the transferred energy. These are the predictions for the test data for which the neural network has not been trained on. 95% confidence interval is shown by the shaded blue area, bounded by red lines.

Image not extracted

Figure 7. The same as Figure 6 but for other kinematical conditions.

yˉ=1N∑i=1Ny^i(12)\bar{y} = \frac{1}{N}\sum_{i=1}^{N}\hat{y}_i \tag*{(12)}
σ2=1N−1∑i=1N(y^i−yˉ)2(13)\sigma^{2} = \frac{1}{N-1}\sum_{i=1}^{N}(\hat{y}_i-\bar{y})^{2} \tag*{(13)}
∣yˉ−μ∣<1.96×σN.(14)\lvert\bar{y}-\mu\rvert< 1.96 \times\frac{\sigma}{\sqrt{N}}. \tag*{(14)}

The estimated average of the predictions is plotted in Figure 6 and Figure 7 together with the estimated 95% confidence interval, bounded by two red lines and shaded in blue. One can observe that the overall agreement of our neural network predictions with the data for 40Ca{}^{40}\mathrm{Ca} remarkable also in the cases not shown.

is satisfactory from quasielastic up to the Δ\Delta resonance. The prediction results could be further improved by adding more examples into the training set, with different choices of the neural-network hyperparameters or including more handcrafted features. One could also use alternative approaches, like Bayesian neural networks to better take into account the uncertainty. However, for the exploratory purpose of the present work, we do not claim optimality in any sense, and our focus is on presenting the interest of the method.

As an additional illustration of the predictive performances of the neural network approach we consider some tests on nuclei of particular interest for the neutrino-oscillation program. Namely oxygen, employed in the present and future water Cherenkov detectors Super-Kamiokande and Hyper-Kamiokande, and argon, the nucleus employed in three detectors of the present short-baseline Fermilab neutrino programs as well as by the future DUNE experiment.

In order to investigate the predictive power of our neural network for some specific kinematical conditions, in the case of oxygen (for which only few data are available, as shown in Table I) we remove from the training data set all the events corresponding to electron scattering on 16O{}^{16}\mathrm{O} with a scattering angle of θ=37.1∘\theta= 37.1^\circ and an incoming electron energy of E=0.737E = 0.737 GeV. The predictions are shown in the left-hand panel of Figure 8. On the same figure, in its middle and right-hand panel, we show the neural network predictions for argon and titanium. For these two nuclei, the inclusive electron-scattering data has been made available only recently [23, 24, 25]. Due to the increasing interest in argon detectors, the results of titanium (Z=22Z = 22) represent an important source of information for the charged current interaction of neutrinos with a neutron of 40Ar{}^{40}\mathrm{Ar} (Z=18Z = 18, N=22N = 22). For 40Ar{}^{40}\mathrm{Ar} and 48Ti{}^{48}\mathrm{Ti}, experimental data only exist for one particular experimental set-up (the same for both nuclei), corresponding to E=2.222E = 2.222 GeV and θ=15.541∘\theta= 15.541^\circ. We removed from the usual dataset data points corresponding to these two nuclei in order to avoid the risk of overfitting due to the presence in the training dataset of the other similar nucleus, with the same kinematical conditions. For all three panels of Figure 8, the represented 95% confidence interval is computed by considering N=40N = 40 independent neural network predictions, following the methodology previously explained for 40Ca{}^{40}\mathrm{Ca} (where it was taken N=100N = 100). In the case of 16O{}^{16}\mathrm{O} the band is very narrow because the confidence intervals are small but in the case of 40Ar{}^{40}\mathrm{Ar} and 48Ti{}^{48}\mathrm{Ti} it is more visible in the figure. We explain this because of the complete absence of 40Ar{}^{40}\mathrm{Ar} and 48Ti{}^{48}\mathrm{Ti} in the training set.

Image not extracted

Figure 8. Comparison between the inclusive (e,e′)(e,e') double differential cross section data and the neural network’s predictions for 16O^{16}\mathrm{O} (left panel), 40Ar^{40}\mathrm{Ar} (middle panel) and 48Ti^{48}\mathrm{Ti} (right panel). The different values of incoming electron energy and scattering angle are written above each panel. For the left panel, the neural network was trained without the data represented on the figure. For the middle and right panels, the neural network was trained without both 40Ar^{40}\mathrm{Ar} and 48Ti^{48}\mathrm{Ti} data. Theoretical predictions of SuSAv2 and GiBUU approaches are also shown. They are taken respectively from Refs. [57] and [14] for 16O^{16}\mathrm{O} and from Refs. [63] and [64] for 40Ar^{40}\mathrm{Ar} and 48Ti^{48}\mathrm{Ti}. GENIE Monte Carlo event generator results, taken from Ref. [20] in the case of 40Ar^{40}\mathrm{Ar} and from Ref. [18] in the case 48Ti^{48}\mathrm{Ti}, are also displayed. The red lines are the boundary of the 95% confidence interval shaded in blue.

For all three cases shown in Figure 8, the neural network predictions are excellent for the whole data set, from the quasielastic bump to the Δ\Delta resonance, passing by the “dip” region. Similar agreement has also been obtained by other theoretical approaches, such as SuSAv2 [57] and GiBUU [65], largely employed in neutrino cross section studies. We briefly recall the reader that SuSAv2, an updated version of the Super Scaling Analysis (SuSA) [66] approach, is based on the observation of the superscaling behavior (i.e. on the simultaneous qq and Fermi momentum independence) of the scaling function (a sort of nuclear response function in terms of a scaling variable) extracted from electron scattering data. Similarly to the neural network, it can be considered as a phenomenological data-driven approach, obviously with a much lower number of parameters and with a sound physical basis which microscopically justify it [67].

The Giessen Boltzmann-Uehling-Uhlenbeck (GiBUU) framework is an implementation of transport theory which allows a description of many nuclear reactions. It allows consistent treatment of the initial reaction vertex and of the final state processes. It takes into account various nuclear effects via the local density approximation for the nuclear ground state, mean-field potentials, and in-medium spectral functions. In the implementation of Ref. [14], from which we take the electron cross section results of oxygen shown in Figure 8, multinucleon excitations were not included, which explains the GiBUU’s underestimation of data in the “dip” region in this case. The underestimation was reduced in the GiBUU calculations for argon and titanium [64] by using an updated version of the model that takes into account the multinucleon contributions. We note that the “dip” region, which is particularly difficult to microscopically describe and reproduce with theoretical approaches, is well predicted by the neural network.

For completeness, in Figure 8 we also plot the results obtained with two different versions of GENIE (one of the most commonly used Monte Carlo event generators in the neutrino community) for 40Ar^{40}\mathrm{Ar} and 48Ti^{48}\mathrm{Ti}, taken from Refs. [20] and [18], respectively. One can observe that starting from the “dip” region GENIE results disagree with the data. This disagreement is analyzed in Refs. [18] and [20], where several results are also shown and discussed for inclusive electron scattering on carbon.

Summary and perspectives

We have deployed a neural network model to predict the electron scattering inclusive double-differential cross sections on nuclei. The neural network predictions have been compared with a large amount of electron scattering data collected in the past as well as with recent data on Argon and Titanium that are of interest to the neutrino community. Our results show that neural networks can reproduce and predict the electron scattering cross section with an accuracy comparable to that provided by the microscopic approaches developed in the past and nowadays generalized to investigate the neutrino-nucleus scattering.

Neural network approaches can be impaired by a limited amount of data. Therefore, it was not a given that the data available was sufficient to train the model to make inference for unseen data. The many orders of magnitude spanned by the cross sections and their varying shape due to the different types of nuclear excitations (quasielastic, multinucleon, resonance excitations, and deep-inelastic) induced in different scattering kinematics make inference challenging. The neural network performance suggests that it could be used as an additional tool in the studies of electron and neutrino scattering on nuclei. Moreover, this can be used to predict the electron cross sections for nuclei and/or for kinematical conditions where experimental data are absent and employ these predictions to validate Monte Carlo simulations. It could also be employed as a support to drive and speed up microscopic evaluation of cross sections and/or response functions, in some sense by generalizing what has been performed in Ref. [43] beyond the quasielastic excitation considered in that work.

Another perspective involves the use of neural networks to directly predict the neutrino-nucleus cross sections. This task is not trivial since the neutrino beams are not monochromatic. So in this case the measured quantity is the flux-integrated double differential cross section in terms of the final state measurable variables. For charged-current scattering process (νl\nu_l, ll) on nuclei, where ll is the charged lepton, this neutrino cross section reads:

d2σνdEldΩ=1∫Φ(Eνl) dEνl∫dEνl[d2σνdωdΩ]ω=Eνl−ElΦ(Eνl),(15)\frac{d^2\sigma_\nu}{dE_l d\Omega} = \frac{1}{\int\Phi(E_{\nu_l})\,dE_{\nu_l}} \int dE_{\nu_l} \left[\frac{d^2\sigma_\nu}{d\omega d\Omega}\right]_{\omega=E_{\nu_l}-E_l} \Phi(E_{\nu_l}), \tag*{(15)}

where ElE_l is the charged lepton energy and dΩd\Omega, the differential solid angle in the direction specified by the charged lepton momentum. They represent the two final state measured variables. This expression formally reduces to that analyzed in this paper, when the flux of incoming particles, Φ\Phi, reduces to a delta distribution. This is the case of monochromatic electron beams. For neutrino scattering, the problem is more complex since for a given set of the measured variables ElE_l and Ω\Omega, one explores the full energy spectrum of neutrinos above the charged lepton energy, being Eνl=El+ωE_{\nu_l} = E_l + \omega. As a consequence, all the reaction channels (giant resonances, quasielastic, multinucleon excitations, pion production arising from nucleon resonances decay, and deep-inelastic scattering) are entangled and isolating a primary vertex process from the measurement of neutrino flux-integrated differential cross section is much more difficult. When employing a deep learning approach for neutrino scattering, the technical challenge would be to design a network sufficiently rich to encode the complexity of the cross section for different primary vertex processes over the phase space relevant to the signal process. A similar approach is currently under study in the high energy collider community [30] in connection with the so called “matrix element method”. This possibility should be investigated with the perspective of the development of an AI-based Monte Carlo event generator for neutrino-nucleus scattering.

Acknowledgements

We would like to thank Fabrice Couderc for his precious suggestions on neural network optimisation. We warmly thank Stefano Fortunati for interesting conversations on statistical topics. We also acknowledge Christophe Coudé, Raphaël Lasseri, Gabriel Perdue, Pierre-Etienne Pion, David Regnier, Guillaume Scamps and Kazuhiro Terao for fruitful discussions and exchanges on neural networks in connection with nuclear and neutrino physics. We thank Gwendolen Rodgers for proofreading the English.

References

References

  1. [16]V. Pandey, N. Jachowicz, T. Van Cuyck, J. Ryckebusch, and M. Martini, Phys. Rev. C92, 024606 (2015), 1412.4624.
  2. [17]G. D. Megias, J. E. Amaro, M. B. Barbaro, J. A. Caballero, and T. W. Donnelly, Phys. Rev. D 94, 013012 (2016), 1603.08396.
  3. [18]A. M. Ankowski and A. Friedland, Phys. Rev. D 102, 053001 (2020), 2006.11944.
  4. [19]J. L. Barrow, S. Gardiner, S. Pastore, M. Betancourt, and J. Carlson, Phys. Rev. D 103, 052001 (2021), 2010.04154.
  5. [20]A. Papadopoulou et al. (electrons for neutrinos), Phys. Rev. D 103, 113003 (2021), 2009.07228.
  6. [21]C. Andreopoulos, A. Bell, D. Bhattacharya, F. Cavanna, J. Dobson, et al., Nucl.Instrum.Meth. A614, 87 (2010), 0905.2517.
  7. [22]M. Khachatryan et al. (CLAS, e4v), Nature 599, 565 (2021).
  8. [23]H. Dai et al. (Jefferson Lab Hall A), Phys. Rev. C 98, 014617 (2018), 1803.01910.
  9. [24]H. Dai et al., Phys. Rev. C 99, 054608 (2019), 1810.10575.
  10. [25]M. Murphy et al., Phys. Rev. C 100, 054606 (2019), 1908.01802.
  11. [26]L. Gu et al. (Jefferson Lab Hall A), Phys. Rev. C 103, 034604 (2021), 2012.11466.
  12. [27]O. Benhar et al. (2014), 1406.4080.
  13. [28]P. A. Machado, O. Palamara, and D. W. Schmitz, Ann. Rev. Nucl. Part. Sci. 69, 363 (2019), 1903.04608.
  14. [29]P. Baldi, P. Sadowski, and D. Whiteson, Nature Commun. 5, 4308 (2014), 1402.4735.
  15. [30]K. Albertsson et al., J. Phys. Conf. Ser. 1085, 022008 (2018), 1807.02876.
  16. [31]A. Radovic, M. Williams, D. Rousseau, M. Kagan, D. Bonacorsi, A. Himmel, A. Aurisano, K. Terao, and T. Wongjirad, Nature 560, 41 (2018).
  17. [32]X. Ju et al., in 33rd Annual Conference on Neural Information Processing Systems (2020), 2003.11603.
  18. [33]P. Calafiura, D. Rousseau, and K. Terao, Artificial Intelligence for High Energy Physics (WORLD SCIENTIFIC, 2022), https://www.worldscientific.com/doi/pdf/10.1142/12200, URL https://www.worldscientific.com/doi/abs/10.1142/12200.
  19. [34]R. Acciarri et al. (MicroBooNE), JINST 12, P03011 (2017), 1611.05531.
  20. [35]G. N. Perdue et al. (MINERvA), JINST 13, P11020 (2018), 1808.08332.
  21. [36]P. Baldi, J. Bian, L. Hertel, and L. Li, Phys. Rev. D 99, 012011 (2019), 1811.04557.
  22. [37]B. Abi et al. (DUNE), Phys. Rev. D 102, 092003 (2020), 2006.15052.
  23. [38]S. Alonso-Monsalve, D. Douqa, C. Jesús-Valls, T. Lux, S. Pina-Otey, F. Sánchez, D. Sgalaberna, and L. H. Whitehead, Phys. Rev. D 103, 032005 (2021), 2009.00688.
  24. [39]P. Abratenko et al. (MicroBooNE), Phys. Rev. D 103, 052012 (2021), 2012.08513.
  25. [40]A. Ghosh et al. (MINERvA) (2021), 2103.06992.
  26. [41]Y. Alanazi, N. Sato, T. Liu, W. Melnitchouk, M. P. Kuchera, E. Pritchard, M. Robertson, R. Strauss, L. Velasco, and Y. Li (2020), 2001.11103.
  27. [42]Y. Alanazi et al. (2020), 2008.03151.
  28. [43]K. Raghavan, P. Balaprakash, A. Lovato, N. Rocco, and S. M. Wild, Phys. Rev. C 103, 035502 (2021), 2010.12703.
  29. [44]L. Alvarez-Ruso, K. M. Graczyk, and E. Saul-Sala, Phys. Rev. C 99, 025204 (2019), 1805.00905.
  30. [45]A. Aguilar-Arevalo et al. (MiniBooNE), Phys.Rev. D81, 092005 (2010), 1002.2680.
  31. [46]S. Nagu, J. Singh, J. Singh, and R. B. Singh (2021), 2105.12733.
  32. [47]P. Bedaque et al., Eur. Phys. J. A 57, 100 (2021).
  33. [48]O. Benhar, D. Day, and I. Sick (2006), nucl-ex/0603032.
  34. [49]O. Benhar, D. day, and I. Sick, Rev. Mod. Phys. 80, 189 (2008), nucl-ex/0603029.
  35. [50]http://discovery.phys.virginia.edu/research/groups/qes-archive/.
  36. [51]M. Martini, M. Ericson, G. Chanfray, and J. Marteau, Phys.Rev. C80, 065501 (2009), 0910.2622.
  37. [52]M. Martini, M. Ericson, and G. Chanfray, Phys.Rev. D85, 093012 (2012), 1202.4745.
  38. [53]M. Martini, M. Ericson, and G. Chanfray, Phys. Rev. D87, 013009 (2013), 1211.1523.
  39. [54]J. Nieves, F. Sanchez, I. Ruiz Simo, and M. Vicente Vacas, Phys.Rev. D85, 113008 (2012), 1204.5404.
  40. [55]O. Lalakulich, U. Mosel, and K. Gallmeister, Phys. Rev. C86, 054606 (2012), 1208.3678.
  41. [56]A. M. Ankowski and C. Mariani, J. Phys. G44, 054001 (2017), 1609.00258.
  42. [57]J. E. Amaro, M. B. Barbaro, J. A. Caballero, R. González-Jiménez, G. D. Megias, and I. Ruiz Simo, J. Phys. G 47, 124001 (2020), 1912.10612.
  43. [58]Goodfellow (2015), URL http://arxiv.org/abs/1512.03385.
  44. [59]K. He, X. Zhang, S. Ren, and J. Sun (2016), URL http://arxiv.org/abs/1512.03385.
  45. [60]M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, et al., TensorFlow: Large-scale machine learning on heterogeneous systems (2015), software available from tensorflow.org, URL https://www.tensorflow.org/.
  46. [61]D. P. Kingma and J. Ba, Adam: A method for stochastic optimization (2014), URL http: //arxiv.org/abs/1412.6980.arxiv.org/abs/1412.6980
  47. [62]I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning (MIT Press, 2016), http://www. deeplearningbook.org.
  48. [63]M. B. Barbaro, J. A. Caballero, A. De Pace, T. W. Donnelly, R. González-Jiménez, and G. D. Megias, Phys. Rev. C 99, 042501 (2019), 1902.06338.
  49. [64]U. Mosel and K. Gallmeister, Phys. Rev. C 99, 064605 (2019), 1811.10637.
  50. [65]O. Buss, T. Gaitanos, K. Gallmeister, H. van Hees, M. Kaskulov, et al., Phys.Rept. 512, 1 (2012), 1106.1344.
  51. [66]J. E. Amaro, M. B. Barbaro, J. A. Caballero, T. W. Donnelly, A. Molinari, and I. Sick, Phys. Rev. C71, 015501 (2005), nucl-th/0409078.
  52. [67]R. González-Jiménez, G. D. Megias, M. B. Barbaro, J. A. Caballero, and T. W. Donnelly, Phys. Rev. C90, 035501 (2014), 1407.8346.

Paper details

vlnc:W_3rjsqbpu
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