Putting machine learning to the test in a quantum many-body system Yilun Gao, Alberto Rodríguez, Rudolf A. Römer Abstract Quantum many-body systems pose a formidable computational challenge due to the exponential growth of their Hilbert space. While machine learning (ML) has shown promise as an alternative paradigm, most applications remain at the proof-of-concept stage, focusing narrowly on energy estimation at the lower end of the spectrum. Here, we push ML beyond this frontier by extensively testing HubbardNet , a deep neural network architecture for the Bose-Hubbard model. Pushing improvements in the optimizer and learning rates, and introducing physics-informed output activations that can resolve extremely small wave-function amplitudes, we achieve ground-state energy errors reduced by orders of magnitude and wave-function fidelities exceeding 99%. We further assess physical relevance by analysing generalized inverse participation ratios and multifractal dimensions for ground and excited states in one and two dimensions, demonstrating that optimized ML models reproduce localization, delocalization, and multifractality trends across the spectrum. Crucially, these qualitative predictions remain robust across four decades of the interaction strength, e.g. spanning across superfluid, Mott-insulating, as well as quantum chaotic regimes. Together, these results suggest ML as a viable qualitative predictor of many-body structure, complementing the quantitative strengths of exact diagonalization and tensor-network methods. I Introduction Machine learning (ML) has recently emerged as a flexible, optimization-driven paradigm for tackling some of the most challenging problems in theoretical and computational physics Alpaydin 2020 ; Hinton and Sejnowski 1999 ; Carleo et al. 2019 ; Mehta et al. 2019 , particularly those involving quantum many-body systems Schollwöck 2011 ; Tasaki 2020 ; Fazio et al. 2025 . These systems exhibit an exponentially growing Hilbert space with system size, making traditional approaches such as exact diagonalization (ED) feasible only for small systems Feynman 1982 ; Troyer and Wiese 2005 ; Jung and Noh 2020 ; Gao and Römer 2025 . While tensor network methods like DMRG Östlund and Rommer 1995 ; Schollwöck 2011 ; Verstraete et al. 2023 and PEPS Orús 2014 ; Schuch et al. 2007 have extended our reach, they too face limitations in higher dimensions or in systems with strong entanglement and disorder Kottmann 2022 . In particular, these methods are not designed for rapid parameter sweeps. Early applications of ML in physics have largely demonstrated proof-of-principle successes, showing that neural-network-based representations of quantum states can capture essential features of ground states Carleo and Troyer 2017 ; Nomura et al. 2017 and, in some cases, outperform conventional variational methods Vivas et al. 2022 ; Kottmann 2022 . Still, systematic tests of wave-function fidelity across distinct interaction regimes are scarce. Beyond proof-of-concept energy and ground state estimates McClean et al. 2016 ; Saito 2017 ; Saito and Kato 2018 ; Hibat-Allah et al. 2020 , a key question remains: Can ML reproduce the structure of many-body eigenstates across phase transitions and deep in the excitation spectrum? This question is central to the connection between eigenstate structure and experimentally relevant observables in regimes governed by quantum chaos and the Eigenstate Thermalization Hypothesis (ETH) Borgonovi et al. 2016 ; D’Alessio et al. 2016 ; Deutsch 2018 ; Trotzky et al. 2012 ; Santos et al. 2012 ; Rigol et al. 2008 . Here we stress-test the recently proposed HubbardNet Zhu et al. 2023 on the Bose-Hubbard model (1D and 2D) Fisher et al. 1989 ; Lewenstein et al. 2007 ; Bloch et al. 2008 ; Bloch et al. 2012 ; Krutitsky 2016 ; Cazalilla et al. 2011 . With optimizer and learning-rate improvements and physics-informed output activations, we achieve sub-percent ground-state energy errors, >99\% wave-function fidelity, and robust predictions of generalized fractal dimensions (GFD) Mirlin 2000 ; Evers and Mirlin 2008 ; Rodriguez et al. 2011 over four decades in the interaction strength parameter. This achievement goes far beyond previous results, showing that a single configuration of a neural network can interpolate between two very distinct physical regimes separated by a quantum phase transition. Furthermore, for excited states, we introduce observable-based training employing GFD that bypasses the need for Gram-Schmidt orthogonalization towers and yields accurate eigenstate-structure statistics in the spectrum bulk across quantum chaos regimes. We thus provide a comprehensive framework to establish how ML-based methods can capture the structural features of many-body states that underpin experimentally relevant observables. While we build on HubbardNet as introduced in Ref. Zhu et al. 2023 , all results concerning multi-order-of-magnitude parameter sweeps, observable-based training of excited states, and activation-function engineering are new to the present work. Our results demonstrate that ML, when carefully optimized, can go well beyond proof-of-concept and offer a viable route to studying complex quantum systems. At the same time, we highlight the challenges that remain—particularly in scaling to larger systems and accurately describing states deep in the spectrum—thus outlining directions for future research. Section II defines the Bose-Hubbard model and structure metrics (the GFD). Section III details HubbardNet and our training improvements. Sections IV - V present ground- and excited-state results, and Sec. VI discusses implications and limitations. II The Bose-Hubbard Model II.1 Definition and notation We study N interacting bosons on a d -dimensional lattice described by the Bose-Hubbard Hamiltonian (BHH) Fisher et al. 1989 , H=-J\sum_{\langle i,j\rangle}\hat{a}_{i}\hat{a}^{\dagger}_{j}+\frac{U}{2}\sum_{i=1}^{M}\hat{n}_{i}(\hat{n}_{i}-1), (1) where \hat{a}_{i}(\hat{a}_{i}^{\dagger}) are annihilation (creation) operators in single particle Wannier states, localized at the different lattice sites i , \hat{n}_{i} being the corresponding number operator, J is the energy associated with particle tunneling between nearest neighboring sites {\langle i,j\rangle} , U represents the onsite pair interaction energy, and M stands for the total number of sites. In the following, we always take the tunneling energy as the energy unit, i.e., J=1 , and periodic boundary conditions are assumed. The BHH models a paradigmatic quantum many-body system Krutitsky 2016 ; Cazalilla et al. 2011 that can be faithfully implemented experimentally using ultracold atoms in optical potentials Lewenstein et al. 2007 ; Bloch et al. 2008 ; Bloch et al. 2012 , and exhibits a rich phenomenology, including a ground state quantum phase transition between a Mott insulator and a superfluid, controlled by the ratio J/U and the bosonic density, which has been experimentally observed Greiner et al. 2002 ; Boéris et al. 2016 ; Bakr et al. 2010 . Additionally, the complex excitation spectrum of the system gives rise to quantum chaos Pausch et al. 2021a ; Pausch et al. 2021b ; Pausch et al. 2022 , inducing the emergence of dynamical ergodic behaviour Dueñas et al. 2025 ; Pausch et al. 2025 . Hence, this bosonic model (and variations thereof) provides an ideal platform for the investigation of strongly correlated dynamics and thermalization in isolated quantum many-particle systems Srednicki 1994 ; Srednicki 1999 ; Borgonovi et al. 2016 ; D’Alessio et al. 2016 ; Deutsch 2018 , questions which are currently being intensively studied both theoretically and experimentally Cheneau et al. 2012 ; Trotzky et al. 2012 ; Ronzheimer et al. 2013 ; Choi et al. 2016 ; Lukin et al. 2019 ; Rispoli et al. 2019 ; Takasu et al. 2020 ; Léonard et al. 2023 . In the one dimensional case (1D, d=1 ), one has a simple chain with M sites, while in 2D ( d=2 ), we assume a square lattice with M=\sqrt{M}\times\sqrt{M} and \sqrt{M} a positive integer. Since Hamiltonian ( 1 ) conserves the total particle number N , the system can be independently diagonalized in Hilbert subspaces of dimension \text{dim}\mathcal{H}=\binom{M+N-1}{N}, (2) which at fixed particle density grows exponentially with N or M [e.g., at unit density \text{dim}\mathcal{H}(N+1)\approx 4\ \text{dim}\mathcal{H}(N) ]. In the cases J=0 (no tunneling) or U=0 (non-interacting), the model is integrable, and the eigenstates can be uniquely identified by M quantum numbers, due to the existence of M observables that commute with the Hamiltonian. For J=0 , the eigenstates are given by the Fock states \ket{f}\equiv\ket{n_{1},n_{2}....,n_{M}}, (3) determined by the sequence of occupation numbers n_{i}\geqslant 0 of the lattice sites such that \sum_{i}n_{i}=N , and where each sequence can be uniquely identified by an integer index f\in[1,\text{dim}\mathcal{H} ]. The latter set of Fock states spans a convenient basis in which to diagonalize the system in the general case with J\neq 0 and U\neq 0 . The k -th eigenstate \ket{\Psi_{k}} , ordered by increasing energy E_{k} , would then be written as \ket{\Psi_{k}}=\sum_{f=1}^{\text{dim}\mathcal{H}}\psi_{k}(f)\ket{f},\quad k=0,1\ldots,\text{dim}\mathcal{H}-1, (4) with corresponding expansion coefficients denoted by \psi_{k}(f) . Our ML outputs will directly target these coefficients. II.2 Characterization of eigenstate structure The two most relevant features of the BHH, namely its ground state phase transition and the emergence of a chaotic phase in its excitation spectrum, can be characterized by the accompanying pronounced changes in the structure of the many-particle eigenstates. The passage from the Mott insulating state to the superfluid as a function of U correlates with a delocalization tendency of the ground state in Fock space that encodes the critical point Lindinger et al. 2019 , which in 1D at unit density is U_{c}\approx 3.3 Kashurnikov et al. 1996 ; Ejima et al. 2011 ; Carrasquilla et al. 2013 ; Boéris et al. 2016 . On the other hand, the chaotic phase is populated by eigenstates that become fully extended (ergodic) in Fock space in the thermodynamic limit Pausch et al. 2021a ; Pausch et al. 2021b ; Pausch et al. 2022 ; Clavero and Rodríguez 2025 . To quantify wave-function structure in the Fock basis, we use the finite-size GFD, D_{q} Mirlin 2000 ; Rodriguez et al. 2011 , which, for \ket{\Psi_{k}} as given above, are defined as D_{q}=\frac{1}{1-q}\frac{\text{ln}\left[\sum_{f}|\psi_{k}(f)|^{2q}\right]}{\text{ln}(\text{dim}\mathcal{H})},\quad q\in\mathbb{R}. (5) The values of D_{q} as \text{dim}\mathcal{H}\to\infty reveal the structural features of the state in the chosen basis for the expansion. In our case, D_{q} probes how wave-function weight is distributed across the Hilbert space: A fully extended state is characterized by D_{q}\to 1 for any q , whereas D_{q>1}\to 0 signals localization in Fock space, i.e., emphasising rare, large-amplitude components. On the other hand, the convergence of D_{q} toward q -dependent values different from 0 and 1 corresponds to multifractality Paladin and Vulpiani 1987 , a property that commonly appears for many-body states in Fock space Lindinger et al. 2019 ; Macé et al. 2019 ; Pausch et al. 2021a and that plays a significant role in disordered systems Evers and Mirlin 2008 ; Rodriguez et al. 2011 ; Pino 2020 . The dimensions D_{q} encode the behaviour of the distribution of wave-function intensities, where different values of q correlate with particular ranges of |\psi_{k}(f)|^{2} , e.g., small wave-function intensities determine D_{q<0} , whereas large intensities contribute most prominently to D_{q>0} . Among all possible D_{q} , we single out D_{1}\equiv\lim_{q\to 1}D_{q}=-\frac{\sum_{f}|\psi_{k}(f)|^{2}\ln|\psi_{k}(f)|^{2}}{\ln(\text{dim}\mathcal{H})}, (6) that converges to the dimension of the support set of the state De Luca et al. 2013 , D_{2} that determines the scaling of the inverse participation ratio \sum_{f}|\psi_{k}(f)|^{4} with Hilbert space dimension, and D_{\infty}=-\frac{\ln\left[\max_{f}|\psi_{k}(f)|^{2}\right]}{\ln(\text{dim}\mathcal{H})}, (7) which is entirely determined by the maximum wave-function intensity. An example of the information conveyed by the GDF is given in Fig. 1 , showing the evolution of D_{1} as a function of U for the entire spectrum of a BHH in 1D (upper panel) and 2D (lower panel) obtained by ED. Figure 1: Variation of the fractal dimension D_{1} (colour scale) as a function of interaction strength U and scaled energy \varepsilon for (top panel) all 1716 states of a 1D BHH with system size M=7 and particle number N=7 , and (bottom panel) a 2D BHH on a 4\times 4 square lattice with particle number N=3 for all 816 states. The grey lines in both panels indicate the upper and lower bounds for the 50 states closest to \varepsilon=0.5 . These trends serve as reference data for the ML predictions analysed in Sections IV and V . The energy spectrum is represented in terms of the scaled energy \varepsilon=\frac{E-E_{0}}{\max_{k}\left\{E_{k}\right\}-E_{0}}, (8) for each value of U . The behaviour exhibited in the upper panel of Fig. 1 is exemplary for all D_{q} Pausch et al. 2021a ; Pausch et al. 2021b and unveils a change in the eigenstate structure associated with the emergence of a chaotic phase: For large interaction, the states are strongly localized in Fock space (low values of D_{1} ), since they become the basis states ( 3 ), and appear organized in degenerate manifolds corresponding to the different eigenenergies of the interaction term in the BHH, E_{k}^{J=0}=(U/2)\sum_{i}n_{i}(n_{i}-1) (for M=N=7 , these amount to 13 different values). As U decreases, the degeneracy is lifted and at U\approx 1 there is a strong basis mixing yielding states that are markedly extended in Fock space ( D_{q}\to 1 ). This region of delocalized states, corresponding to the chaotic phase, extends into the low interaction regime. While the existence of quantum chaos in the 2D BHH has not yet been properly investigated (due mainly to the very demanding numerical analysis), the behaviour of D_{1} in the lower panel of Fig. 1 , for a 4\times 4 system with particle number N=3 , also reveals an intermediate regime around U\approx 2 in the bulk of the spectrum populated by fairly delocalized states in Fock space. In this case, however, since particle density is much lower than unity ( N/M\simeq 0.19 ), 69\% of the eigenstates converge to a non-interacting configuration in the limit U\gg 1 (i.e., with vanishing energy in that limit). Consequently, such highly degenerate manifold mixes a very large number of basis Fock states very quickly as U decreases, giving rise to delocalized states much earlier as compared to the 1D case above. III The ML approach to the BHH At its core, ML is an optimization paradigm—an approach that aligns naturally with the variational principles of quantum mechanics Schrödinger 1926 . Essentially, ML optimizes the coefficients of many-body states, i.e. when expressed in a Fock basis Carleo and Troyer 2017 . Several neural network architectures have been proposed to represent many-body quantum states, including FermiNet Spencer et al. 2021 , PauliNet Hermann et al. 2020 , BosonNet Denis and Carleo 2025 , and HubbardNet Zhu et al. 2023 . We choose HubbardNet as a deliberately minimal, easily reproducible deep neural-network baseline—simple enough to implement (from the explicit code base on its github repository Zhu et al. 2023 ) and audit, yet expressive enough to support a rigorous, end-to-end evaluation beyond energies. III.1 The principles of HubbardNet HubbardNet is a fully-connected multilayer perceptron (MLP) Rumelhart et al. 1986 design as shown schematically in Fig. 2 : Inputs are the occupation numbers \{n_{i}\}_{i=1}^{M} , the interaction strength U , and the particle number N . The neural network (NN) contains four hidden layers ( 400 neurons, \tanh ) and produces a single real output u_{k}(f) 1 1 1 Zhu et al. 2023 use two neurons in the output layer corresponding to the real and imaginary parts of the wave-function. As the wave-functions for the BHH, with a real hopping constant, and without gauge fields, can always be chosen to be real, the use of a single output neuron is sufficient for our purposes. . The latter output is converted into the wave-function coefficient associated with Fock state \ket{f} [see Eqs. ( 3 ) and ( 4 )] via an activation function \sigma(u) plus normalization, \psi^{\textrm{NN}}_{k}(f)=\frac{\sigma[u_{k}(f)]}{\|\sigma[u_{k}]\|},\quad f=1,\ldots,\text{dim}\mathcal{H}, (9) where the superscript NN denotes neural-network prediction. n_{1} n_{2} n_{3} n_{4} U N u_{k}(f) Figure 2: The MLP structure of HubbardNet with M=4 . The green circles represent the input layer with the n_{i} ’s, U , and N as the input parameters, while the purple circles denote the 4 hidden layers. The output neuron is indicated by the red circle. By default, HubbardNet minimizes the following sum of Rayleigh quotients, \mathcal{L}_{k}=\sum_{\gamma}\left(\frac{\bra{\Psi_{k}^{\text{NN}}}H_{\gamma}\ket{\Psi_{k}^{\text{NN}}}}{\braket{\Psi_{k}^{\text{NN}}|\Psi_{k}^{\text{NN}}}}-\mathcal{L}_{P}\right), (10) where \gamma stands for a collection of training points (e.g., multiple values of U and/or of N ), and \mathcal{L}_{P} represents an orthogonality penalty used only for excited states ( k>0 ) to discourage overlap with previously learned lower states Zhu et al. 2023 . The MLP employs an ingenious cosine annealing scheme Loshchilov and Hutter 2017 with a resetting period of 1000 steps in order to avoid getting stuck in local minima. For the ground state \ket{\Psi_{0}} , HubbardNet uses an exponential activation function \sigma\left(u\right)=\exp\left(u\right) in the output layer since the expansion coefficients \psi_{0}(f) are always positive in the BHH. For excited states, Zhu et al. 2023 choose a linear activation function, \sigma(u)=u , and find the desired target excited state \ket{\Psi_{k}} by iteratively constructing the full tower of lower-lying states \ket{\Psi_{k^{\prime}}} with k^{\prime}=0,1,\ldots k-1 and employing a Gram-Schmidt process to ensure orthogonality of the eigenspace. The computational bottleneck is the per-iteration Rayleigh quotient, which scales with the number of sampled Fock configurations. For M=N=7 ( \text{dim}\mathcal{H}=1716 ), energies via ED are seconds-scale; ML requires \gtrsim 10^{4} iterations but amortizes across U once trained. With this setup, Zhu et al. 2023 showed promising energy and state predictions for U\in[1,15] . We demonstrate robust generalization across four decades in U , high-fidelity wave-functions, and accurate eigenstate-structure metrics—and we introduce observable-based training for excited states that avoids Gram-Schmidt towers. III.2 Putting HubbardNet to the test III.2.1 Training & out-of-data predictions We choose the training set of U to extend from the weakly interacting regime, starting with U=0.01 , to the strongly repulsive regime at U=100 . This very large range of U , spanning 4 orders of magnitude, allows us to fully test the convergence and stability of HubbardNet . In order not to prejudice the training of the NN to either regime, we choose 9 training U values equally spaced on a \log -scale, \log_{10}U^{\text{(train)}}=\left\{\pm 2,\pm 1.5,\pm 1,\pm 0.5,0\right\}. (11) To make the training process converge across this wide range of U values, we reduce the maximum learning rate from 0.01 , as used by Zhu et al. 2023 , to 5\times 10^{-6} . We also find that changing the optimizer from stochastic gradient descent to the adaptive momentum method implemented in Adam Kingma and Ba 2017 enables the NN to yield a prediction of E_{0} being \approx 100 times more accurate. Furthermore, we observe that the training does not converge if a constant learning rate is utilized. Therefore, we continue to apply HubbardNet ’s cosine-annealing scheme in the training process, resetting the learning rate to 5\times 10^{-6} every 1000 epochs. Figure 3 gives a typical learning curve of training for 1D ground state energy, i.e., minimizing \mathcal{L}_{0} . We observe that the loss function oscillates dozens of times before reaching a minimum due to the application of the cosine-annealing scheme. In our case, convergence is reached when the standard deviation of the last 50 iterations falls below 10^{-7} . Figure 3: Example of a typical learning curve, indicating the loss function minus its minimum value (reached at convergence), \Delta\mathcal{L} , as a function of training steps for energy-based training of the 1D ground state for all U values with M=N=7 . Once the training is complete, we proceed to use the NN to predict energy and state features at 69 “out-of-data” Arora and Zhang 2021 U values equally spaced on log scale in the interval \log_{10}U\in[-2.1250,2.1250] , corresponding to 0.0075\leqslant U\leqslant 133 . Unless otherwise stated, the training and test U -sets will be the same for all the studies presented in the manuscript. Regarding the 1D ground state, we shall be interested in discussing how the NN performs in the superfluid phase, close to the transition, and in the Mott phase. Here we choose U=0.2054 ( \approx 0.21 ), 2.054 ( \approx 2.05 ) and 20.54 , corresponding to mid-points between adjacent training values, to be representative for the three regions. The latter U values will also be taken as a reference in 2D and when analysing the excitation spectrum. III.2.2 Physics-based improvements While additional ML hyperparameter tuning—beyond learning rate—(e.g., depth/width of layers, regularization, batch size, optimizer schedules) can further improve numerical performance, we deliberately avoid a purely ML-centric optimization loop. Instead, we focus on physics-based refinements that directly influence what the network must represent, aiming at (i) working with an apt loss function, and (ii) selecting an optimal output activation. In particular, these considerations play a most significant role in the study of the system’s excitation spectrum. As we show later, a careful choice of output activation turns out to be central: Given that we intend to have no explicit control over the bounds for the output variable u_{k} (which acquires values in [-17,12] in our 1D simulations), the function \sigma(u) directly controls how well the NN can resolve very small wave-function amplitudes. In combination with a loss function that targets the many-particle state structure via the information encoded in the GDF, we enable the NN to successfully describe the evolution of the eigenstate-structure statistics deep in the spectrum bulk as a function of the interaction strength, accounting for the emergence of a quantum chaos regime. Prioritizing such physics-informed choices keeps the model interpretable and ensures that observed gains reflect improved physical resolution rather than purely algorithmic tuning. IV Ground state properties IV.1 The ground state for the 1D chain We focus on the BHH at unit density with M=N=7 . This is the physically most interesting case in which the system, in the thermodynamic limit, exhibits a transition from a superfluid to a Mott insulator phase at U_{c}\approx 3.3 Kashurnikov et al. 1996 ; Ejima et al. 2011 ; Carrasquilla et al. 2013 ; Boéris et al. 2016 . Here, we consider the loss function in Eq. ( 10 ), and our choice for the activation funcion is \sigma(u)=\exp(u) , as considered by Zhu et al. 2023 . IV.1.1 Ground state energies & fidelities In Fig. 4 we present the results of HubbardNet when trained and tested as explained in Sec. III.2 . Figure 4: Accuracy of the energy-based ground state training in 1D: The upper panel shows E_{0} for M=N=7 as a function of interaction U in terms of the NN (blue \times ) and the ED (solid black line). The \exp(u) label indicates that the exponential \sigma has been used (see text). The middle panel gives \delta E_{0} [Eq. ( 12 )] with the horizontal dashed line indicating a 10^{-2}\equiv 1\% value. The infidelity 1-\mathcal{F} is shown in the lower panel. The nine vertical black dashed lines indicate the training values of U [Eq. ( 11 )]. The upper panel of Fig. 4 compares E_{0} obtained via NN and ED. Overall, E_{0}^{\text{NN}} matches extremely well with E_{0}^{\text{ED}} across the whole range of U values, up to minor deviations for U\gtrsim 10 . In the middle panel of Fig. 4 , the relative error \delta E_{0}=\left|\frac{E^{\text{ED}}_{0}-E^{\text{NN}}_{0}}{E^{\text{ED}}_{0}}\right| (12) is plotted against U . We first note that, as expected, \delta E_{0} for all 9 of the training U values is <1\% . This high level of accuracy remains for test U values up to U\lesssim 5 . As U increases further, the relative deviation \delta E_{0} becomes larger due to the fact that the exact E_{0} progressively approaches zero (the exact ground state energy for U=\infty ) but remains small in absolute terms. In the lower panel of Fig. 4 , we show the state infidelity 1-\mathcal{F} as a function of U , where the fidelity \mathcal{F} is defined as \mathcal{F}=\braket{\Psi_{0}^{\text{ED}}|\Psi_{0}^{\text{NN}}}. (13) A smaller value of infidelity corresponds to a better agreement between \ket{\Psi_{k}^{\text{ED}}} and \ket{\Psi_{k}^{\text{NN}}} . We see that the infidelity remains below \sim 1\% over the whole U range except for the largest out-of-data U value. This is a remarkable result: the NN, slightly modified from the original HubbardNet set-up in terms of improved learning rate and better ML optimizer, can now faithfully reproduce the ED results for E_{0} and produce a good overlap with \ket{\Psi_{0}} . Training only for 9 values of U yields a NN configuration that can provide a full and reliable parameter sweep over four decades of the interaction strength, and that is available to users of the BHH. IV.1.2 Fock basis coefficients and ground-state-structure statistics We now turn to investigating the agreement of individual Fock state coefficients \psi_{0}(f) between \ket{\Psi^{\text{ED}}_{0}} and \ket{\Psi^{\text{NN}}_{0}} . In Fig. 5 (a), we plot the sequence \psi_{0}(f) as a funcion of the index f at the values U=0.21 , 2.05 , and 20.54 mentioned above. (a) (b) Figure 5: Wave function properties from energy-based ground state training in 1D (cp. Fig. 4 ): (a) Coefficients \psi_{0}(f) and (b) distribution P(\alpha) obtained via the NN ( \times , red) and the ED (black lines) at U=0.21 (top), U=2.05 (middle) and U=20.54 (bottom) for M=N=7 . The horizontal Fock-space index in (a) goes from 1 to \text{dim}\mathcal{H}=1716 in integer steps. Lower and upper horizontal axes in each (b) panel show the equivalent values of \alpha and |\psi|^{2} , respectively. For all three U values across the emerging superfluid and Mott insulator regimes—entailing a widely varying state structure, the amplitudes of the wave-function can be captured very well. To be specific, at U=0.21 , the NN predicts correctly \psi_{0}(f) for nearly each Fock index f . As U increases to U=2.05 , the NN can still determine the positions of the larger amplitudes, with very small deviations becoming visible. At U=20.54 , the NN reproduces well the dominant contribution of the homogeneous Fock state \ket{1,1,\ldots,1} at f=1079 and also identifies the larger subleading contributions from other basis states with minor fluctuations. A more complete picture of the predictive power of the NN concerning the state structure can be obtained by analysing the ditribution of wavefuncion intensities. We then construct the histograms {P}(\alpha) of singularity strengths \alpha , defined as Rodriguez et al. 2009 \alpha(f)\equiv-\frac{\text{ln}|\psi(f)|^{2}}{\text{ln}(\text{dim}\mathcal{H})}\geqslant 0. (14) One can see from ( 14 ) that small (large) intensities |\psi(f)|^{2} yield large (small) \alpha values: |\psi|^{2}\to 1\Rightarrow\alpha\to 0 , and |\psi|^{2}\to 0\Rightarrow\alpha\to\infty . We compare the resulting ED and NN distributions in Fig. 5 (b). For U=0.21 and 2.05 , we see an excellent agreement, and the NN reproduces remarkably well the state amplitudes across their entire range, up to \alpha\approx 4 , corresponding to |\psi_{0}(f)|>10^{-7} . However, at U=20.54 , the NN result for P(\alpha) , shown in the bottom panel of Fig. 5 (b), is quite different from the ED result. Despite the appealing visual agreement of the amplitudes discussed earlier, the NN fails to describe correctly the set of very small wave-function intensities. In fact, P^{\text{NN}}(\alpha) exhibits a sharp cut-off at \alpha=4.6 , and hence does not capture amplitudes below 4\times 10^{-8} , whereas the exact distribution extends beyond \alpha=9 , corresponding to |\psi_{0}|<10^{-15} . It must be noted that this limitation is in fact a consequence of the chosen exponential activation function: The observed range for the u output of the NN in this case ranges from -17.02 to 0.097 . Therefore, according to Eq. ( 9 ) with \sigma(u)=\exp(u) , one can conclude that amplitudes below 4\times 10^{-8} cannot be resolved, as observed. The upper (lower) bound for \alpha ( |\psi_{0}| ) imposed by the output activation function is then responsible for the spurious accumulation of P^{\text{NN}}(\alpha) observed around \alpha=4 . An appropriatly adjusted choice of \sigma(u) , however, could circumvent the limitation to sample small wave-function intensities, as we show later on. IV.1.3 Predicting the GDF The quality of the NN prediction regarding the state structure can also be conveniently quantified by evaluating the GDF D_{q} introduced in Section II.2 . Figure 6: Estimates of D_{1} and D_{\infty} versus U from the energy-based ground state training, in 1D with M=N=7 and 2D on a 4\times 4 square lattice with N=3 : For the 1D case, the red (green) \times represent D_{1} ( D_{\infty} ) obtained via the NN, whereas solid curves with the same color are the results from ED. In 2D, the blue (orange) + symbols indicate D_{1} ( D_{\infty} ) from the NN; solid curves of the same color show ED results. In Fig. 6 , we plot, e.g., the values of D_{1} and D_{\infty} against U calculated from the NN and ED. Both D_{1}^{\text{NN}} and D_{\infty}^{\text{NN}} show a remarkable agreement with D_{1}^{\text{ED}} and D_{\infty}^{\text{ED}} across the whole range of U values, and only small deviations become visible as the strong interacting regime is approached ( U\gtrsim 10 ), in accordance with the behaviour observed for E_{0} in Fig. 4 . We see results of similar quality for other D_{q} with q=0.5,2 and 3 (not shown). In particular, note how the evolution of D_{q} from high values for weak interaction toward zero for large U reveals the dramatic change in the state structure, from being large delocalized to becoming strongly localized in Fock space (manifesting the emergent superfluid to Mott insulator transition Lindinger et al. 2019 ), a physically relevant development that the NN has no issues accounting for in its configuration. This result is very encouraging. Even when simply training to reproduce E_{0} , the small deviations observed in the dominant \psi^{\text{NN}}_{0}(f) for U\gtrsim U_{c} do not necessarily lead to enhanced disagreements in these D_{q} . On the other hand, for q<0 , we find that the predictions by the NN are considerably worse, with relative errors above 1\% and noticeable deviations for large interactions. This is not surprising, given that negative q values rely on very small wave-function amplitudes, which the NN struggles to reproduce faithfully, as discussed above and observed in the bottom panel Fig. 5 (b). We emphasize, nonetheless, that a solution to the latter problem is in principle at hand via an appropriate choice of the output activation function. IV.2 The ground state in 2D Figure 7: Accuracy of the energy-based ground state training in 2D: The upper panel shows E_{0}(U) in terms of the NN (blue + ) and the ED (solid black line) of a 4\times 4 Bose-Hubbard square lattice with particle number N=3 . The middle panel gives the relative error \delta E_{0} while the infidelity 1-\mathcal{F} is shown in the lower panel. The vertical black dashed lines and the single horizontal dashed line are as in Fig. 4 . (a) (b) Figure 8: Wave function properties from energy-based ground state training in 2D (cp. Fig. 7 ): (a) Coefficients \psi_{0}(f) and (b) distributions P(\alpha) obtained via the NN ( + , red) and ED (black lines) for a 4\times 4 square lattice with N=3 ( f=1,2,\ldots,816 ) using the same three test U values as in Fig. 5 . Moving towards the 2D case, we choose system size M=16 in a 4\times 4 square with particle number N=3 , and Hilbert space dimension \text{dim}\mathcal{H}=816 , as also considered by Zhu et al. 2023 . This gives us an opportunity to assess the performance of the NN at non-integer fillings. Note, however, that the ground state of this system does not undergo a U -driven phase transition in the thermodynamic limit. As in the 1D case, we choose an energy-based training with an exponential output activation. We show E_{0} , \delta E_{0} and 1-\mathcal{F} as obtained from the NN and ED in Fig. 7 . We find that the NN performs with \delta E_{0}<1\% up to U\approx 50 , even better than in the 1D case. The prediction of the ground state, as characterized by 1-\mathcal{F} , appears to be of similar quality as for 1D. In Fig. 8 (a), we compare the wave-function amplitudes \psi_{0}^{\text{NN}}(f) and \psi_{0}^{\text{ED}}(f) . We find an excellent match overall for the three reference U values shown, and the NN only seems to underestimate some components with large amplitudes in the strong interacting regime, U=20.54 . Similarly, the prediction for the distributions P(\alpha) shows a very good agreement with the exact results. In particular, note that for large U the prediction is much better than in the 1D case as the state does not exhibit amplitudes |\psi_{0}(f)|<10^{-3} . Since we are not working at integer density, there is no emergent superfluid to Mott insulator phase transition, which means that the ground state structure does not change as drastically as we have seen in Fig. 5 (a). Therefore, in this case, it should be easier for the NN to accommodate the dependence of the state features on the interaction strength. The NN assessment of the state structure via the evolution of the fractal dimensions D_{1} and D_{\infty} as functions of U is shown in Fig. 6 . The match with the exact ED results is excellent up to the value U\approx 20 , beyond which the discrepancies become more noticeable, arguably due to the cumulative effect of the small deviations observed in the dominant wave-function amplitudes for strong interactions. Nonetheless, the global tendency is correctly described, and in contradistinction to the 1D case, the lack of an emergent Mott phase in our 2D system allows the ground state to remain largely delocalized over the Fock basis, as reflected by the observed high values of D_{1} across the entire U range. Overall, we find a remarkable performance of the NN in 2D, similar to the 1D results, where the ground state energy and structural features can be fairly well reproduced across a wide range of interaction strength spanning four orders of magnitude. V Excitation spectrum of the BHH In the BHH, the excited states deep in the bulk of the spectrum—those far above the ground state—are crucial for understanding the emergence of quantum chaos and the onset of ergodicity Kolovsky and Buchleitner 2004 ; Pausch et al. 2021a . More generally, excited states in quantum many-body systems govern the physics at finite temperatures and may obey the Eigenstate Thermalization Hypothesis (ETH) Rigol et al. 2008 suggesting that individual eigenstates can mimic thermal ensembles for local observables, bridging microscopic quantum mechanics with macroscopic thermodynamics. Moreover, phenomena like many-body localization (MBL) Yao and Zakrzewski 2020 ; Chen et al. 2025 occur in these excited states, where, upon inclusion of disorder, thermalization may be prevented even at high energy densities—an effect with implications for quantum information storage. It is therefore encouraging to see that Zhu et al. 2023 have also proposed a way to construct a particular target excited state \ket{\Psi_{k}} . Their method involves an iterative construction, via the NN, of the full, orthogonalized, tower of lower-lying excited states, \{\ket{\Psi_{l}}\} for l2 , though the agreement with the ED results in this region is clearly better when using the exponential activation function. The focus, however, must be put on the capability of the NN to reconstruct the wave-function intensity distributions that we show in Fig. 9 (a) for U=0.21,2.05 and 20.54 . For \sigma(u)=\exp(u) (cyan histograms), we see an excellent description of the exact distribution for the lowest interaction, and a reasonable overlap at U=2.05 . The very good performance of the NN for weak interaction strengths up to U\approx 1 is confirmed by the very low values of d_{\text{KL}} observed in panel (d). In this very same U range, the NN trained with the linear activation function performs noticeably worse, as indicated by the larger d_{\text{KL}} values in (d), and visually evident in (a) (magenta histograms). Most importantly, the NN fails very significantly to describe correctly the high interaction regime, U>3 , for either \sigma(u) , manifested by the obvious deviations of P^{\textrm{NN}}(\alpha) from the ED result (black line) for U=20.54 , and the marked surge in the Kullback-Leibler divergence in panel (d). The poor agreement of the distributions for large U could be partially expected, since the training only included GFD for positive q values, which mainly encode the behaviour of the larger wave-function intensities. A correct description of the many-body eigenstate structure for strong interactions hinges on tracking the behaviour of the small state intensities. Then, on the one hand, we need to feed such information at the training stage by considering also GFD with negative q , and secondly, one must overcome the limitations of the prior activation functions to yield small Fock coefficients, as discussed for the ground state in Sec. IV.1.2 . Hence, to increase the coverage of wave-function amplitudes, while maintaining the essential monotonicity of the u\leftrightarrow\psi(f) mapping, we propose to use the activation function \sigma(u)=\exp[\text{sgn}(u)u^{2}] (19) in combination with the training set q_{-}^{\textrm{(train)}}=\{-2,-1,1,2,\infty\}. (20) The ensuing prediction for P(\alpha) of this approach is shown by the green histograms in Fig. 9 (a), and the green points for d_{\text{KL}} in panel (d). The NN captures remarkably well the distribution at U=20.54 , including wave-function amplitudes down to 10^{-15} . The performance of the NN configuration in this case is excellent for U>3 as indicated by the near-zero values of d_{\text{KL}} , while maintaining a very reasonable output also for low interactions. Overall, the latter training strategy yields the most well behaved and balanced prediction for the eigenstate structure of excited states across the whole interaction range. For completion, panel (c) shows the corresponding prediction for D_{1} and D_{-1} versus U . To determine if our approach remains valid at other spectral regions, we repeat the analysis for scaled energy \varepsilon=0.25 . The results obtained for the three activation functions are presented in Fig. 10 . The conclusions are qualitatively the same as for the previous energy. The NN reproduces correctly the overall dependence of the GFD on interaction strength, capturing the pronounced change around U\approx 5 associated with the border of the chaotic regime. The agreement with ED is very good within the chaotic phase for positive q while larger relative errors \delta D_{q} emerge for strong interactions. When it comes to the P(\alpha) distributions, the linear activation function performs very poorly for any U , as observed in the (a) and (d) panels, and while \sigma(u)=\exp(u) works fairly well for weak and moderate interactions, it clearly fails for U>5 . As before, the training based on Eq. ( 19 ) and q_{-}^{\textrm{(train)}} produces the NN with the best performance throughout the entire U range, albeit its accuracy seems arguably lower than for \varepsilon=0.5 . We have thus demonstrated that a clever D_{q} -based training combined with an appropriate choice of activation function leads to a NN configuration capable of describing correctly the physically relevant changes in the many-body eigenstate structure deep in the bulk of the spectrum across a substantial interaction range, without the need to target the precise Fock space expansion of selected excited states. Our activation function in Eq. ( 19 ) is guided by physical intuition but there might be even more efficient choices. V.3 Excited states for the 2D case (a) (b) (c) (d) (e) Figure 11: D_{q} -based excited state training for the 2D BHH on a 4\times 4 square lattice with N=3 , at different scaled energies, and for linear and exponential activation functions with q_{+}^{\textrm{(train)}} . Panels (a) and (b) give the distributions P(\alpha) at \varepsilon=0.5 and 0.25 , respectively, for the indicated U values. Panels (c) and (d) show D_{1} , D_{\infty} versus U at \varepsilon=0.5 and 0.25 , respectively. Panel (d) displays d_{\text{KL}}(P^{\textrm{NN}},P^{\textrm{ED}}) at the two scaled energies considered. The use of color bars, lines and symbols is the same as in Figs. 9 and 10 . We also illustrate D_{q} -training for excited states in the 2D BHH at non-integer filling, considering the 4\times 4 square lattice with N=3 particles, as we did for the ground state analysis in Sec. IV.2 . The NN results using linear and exponential activation functions with q_{+}^{\textrm{(train)}} are shown in Fig. 11 for scaled energies \varepsilon=0.5 and \varepsilon=0.25 . The predicted GFD in panels (c) and (d) reproduce fairly well the dependence of the exact D_{1} and D_{\infty} on U , with slightly larger fluctuations for strong interaction as is to be expected, but arguably even more accurately that in the 1D case. The evolution of the GFD reveals that the change in the excited eigenstate structure as a function of U in 2D is more gentle than in the 1D chain. Accordingly, we see in panels (a) and (b) that the domain of the wave-function intensity distributions does not change significantly with U , keeping \alpha\leqslant 3 ( |\psi(f)|\geqslant 4\times 10^{-5} ) even at U=20.54 . The performance of the NN when predicting the P(\alpha) distribution as a function of U may seem only moderate, better for exponential activation, but note that the values of the KL divergence stay consistently below 0.5 up to U\approx 10 , which is in agreement with the behaviour observed in 1D. Despite the absence of very small wave-function intensities, the distinct increase of KL for large interaction suggests that the NN training would similarly benefit from the use of an activation function with an enhanced sensitivity for small Fock coefficients. VI Discussions and Conclusions Let us briefly summarize what we have achieved thus far. We utilized a dense neural network (NN) ( HubbardNet Zhu et al. 2023 ) to study interacting quantum many-body systems, with a focus on the Bose-Hubbard Hamiltonian (BHH). By just slightly improving on the learning rate and the optimizer—but without changing the structure of the NN—and training with respect to the energy, we show that HubbardNet is capable of predicting the ground state energy E_{0} , as well as the wave-function \ket{\Psi_{0}} both in 1D and 2D across four orders of magnitude of the interaction strength U . This means that one can get access, when given such a trained NN for the system sizes and densities considered, to the information of E_{0} and \ket{\Psi_{0}} at any interaction within the range U\in[0.01,100] without the need to use any other numerical method, such as exact diagonalization. We emphasize that this includes U values corresponding to states with completely different features. The NN furthermore enables us to capture important physical observables calculated from \ket{\Psi_{0}} . We showed as an example that the fractal dimensions D_{q} and the wave-function intensity distribution can be well characterized across the whole U range. Excited-state training is notably more difficult. We find that training directly on average physical observables, i.e., the fractal dimensions D_{q} 2 2 2 Although the GFD are not quantum mechanical observables in the strict sense, they encode key structural information about the states that ultimately determine observable expectation values, especially in the vicinity of critical points. Hence, we anticipate that training on actual observables would lead to results comparable in predictive strength. , offers a distinct advantage. This method, coupled with a physically motivated choice of the activation function, facilitates qualitatively accurate spectrum-wide predictions of the many-body state structure, as tested for scaled energies \varepsilon=0.25 and 0.5 , while bypassing the complexities of Gram-Schmidt towers. Ultimately, this provides a more scalable framework for estimating physical quantities without the overhead of explicit wave-function reconstruction. In particular, we have successfully used this approach to describe the evolution of the excited state structure with U across a chaotic phase. Looking toward the future, our results suggest a natural and distinct role for today’s NN-based machine learning (ML) in quantum many-body physics. At present and for the as of today still relatively small system sizes accessible to NN approaches, ML methods are unlikely to be competitive—at least in the near term—with highly specialized and rigorously optimized techniques such as tensor-network algorithms Östlund and Rommer 1995 ; Schollwöck 2011 ; Verstraete et al. 2023 ; Orús 2014 ; Schuch et al. 2007 or modern implementations of exact diagonalization Troyer and Wiese 2005 ; Jung and Noh 2020 ; Weinberg and Bukov 2019 ; Bollhöfer and Notay 2007 ; Balay et al. 1997 ; Balay et al. 2019 ; Hernandez et al. 2005 ; Roman et al. 2017 ; Pietracaprina et al. 2018 ; Nataf and Mila 2014 ; Dabholkar and Alet 2024 . These established methods are likely to remain indispensable for high-precision, large system size, benchmark-quality calculations. However, as we have demonstrated here, current ML approaches offer a complementary strength: They can provide rapid and simultaneous access to energy and wave-function information of a medium-sized quantum many-body system across a broad range of physically relevant continuous parameters. Once trained, an NN effectively encodes this information in a compact form, enabling dense parameter sweeps over interaction strength, particle density, or other control parameters at negligible additional computational cost. This capability opens a valuable use case. ML models can serve as an exploratory tool, allowing newcomers and experienced practitioners alike to obtain a quick, global overview of the bandwidth, phase structure, and qualitative physics exhibited by classes of quantum many-body systems, such as, e.g., Heisenberg, Hubbard, tJ , Haldane-Shastry and related models, without committing extensive resources upfront. Armed with this broad overview, as encoded in trained NNs, researchers can then identify physically interesting regimes and formulate more targeted questions. These focused problems can subsequently be addressed using bespoke, highly optimized numerical methods, such as DMRG/PEPS, that excel in restricted regions of parameter space. In this way, ML might not replace traditional approaches but instead act as a front-end guide, streamlining and sharpening the overall workflow of quantum many-body investigations. Acknowledgements. We gratefully acknowledge funding by the University of Warwick’s International Partnership Fund 2024 and support from Warwick’s Scientific Computing Research Technology Platform (RTP) for its high-performance computing facilities and the Sulis Tier 2 HPC platform hosted by the RTP. Sulis is funded by EPSRC Grant EP/T022108/1 and the HPC Midlands+ consortium. A.R. acknowledges support through grant no. PID2024-156340NB-I00 funded by Ministerio de Ciencia, Innovación y Universidades/Agencia Estatal de Investigación (MICIU/AEI/10.13039/501100011033) and by the European Regional Development Fund (ERDF). UK research data statement: ML codes for the adapted HubbardNet are available at https://github.com/DisQS/HubbardNet/tree/main .