Neural-network quantum states for solving few-body problems: application to Efimov physics Sora Yokoi, Shimpei Endo, Hiroki Saito Abstract Neural-network quantum states have recently emerged as a powerful method for solving quantum many-body problems, with notable successes in lattice systems. Here, we extend this approach to strongly interacting few-body problems in continuous space, and demonstrate its capability by computing the Efimov states and associated few-body bound states. Using a fully connected feedforward neural network with Jacobi coordinates as inputs, combined with a projection method, we compute the ground and first excited states for three- to six-body systems of identical bosons at unitarity, as well as a mass-imbalanced fermionic system consisting of two identical fermions and a third particle. The obtained energies of the ground and first excited states agree well with previously reported results. Furthermore, the proposed approach also reproduces key features of Efimov states, including the discrete scale invariance, the characteristic geometric structure of the wave function, and the critical-mass behavior in mass-imbalanced fermionic systems. Our method can be readily applied to a broad class of strongly correlated few-body problems in continuous space. I Introduction Numerical calculations for quantum many-body systems face fundamental limitations due to the exponential growth of the Hilbert space with an increase in the degrees of freedom. To overcome this problem, diverse numerical approaches have been developed, including quantum Monte Carlo methods QMC , density-matrix renormalization groups DMRG , and tensor networks TN . In addition to these approaches, novel methods based on artificial neural networks have recently attracted much interest NQS ; review1 ; review2 ; review3 ; review4 . Leveraging the flexibility and adaptability of neural networks for representing various features, many-body wave functions are encoded on neural networks with a comparatively small number of network parameters, which are optimized using machine learning techniques. The method of neural-network quantum states was first applied to spatially discrete problems, such as spins and particles on lattices NQS ; BHM ; Nomura1 ; Cai ; BHM2 , and then extended to continuous-space problems, such as interacting bosons boson1 ; boson2 ; boson3 ; boson4 ; boson5 , fermionic systems fermi1 ; fermi2 ; fermi3 ; fermi4 ; fermi5 ; fermi6 ; fermi7 ; fermi8 , and nucleons in nuclei nuclei1 ; nuclei2 ; nuclei3 ; nuclei4 ; nuclei5 ; nuclei6 ; nuclei7 . Neural-network quantum states are not only powerful tools for ground states, but also useful for exploring excited states Choo ; Nomura ; Viteritti ; Entwistle ; Pfau ; Li . An excited state belonging to a symmetry sector different from the ground state can be simply obtained by minimizing the energy by fixing quantum numbers, such as momentum or magnetization. Even if the desired excited state cannot be distinguished from the ground state by quantum numbers, one can still compute the excited state by projecting it to the Hilbert space orthogonal to the ground state. Using these procedures, low-lying excitations of the Heisenberg model, the Bose-Hubbard model Choo , and the J_{1} - J_{2} model Nomura ; Viteritti have been calculated. More recently, multiple excited states of benzene-scale molecules have been computed by regarding them as the ground state of an extended system Pfau . While neural-network quantum states have been successfully applied to a variety of quantum systems, it remains unclear to what extent they can serve as viable tools for strongly correlated quantum few-body problems, which are of pivotal importance in nuclear physics and cold atomic systems. In nuclear physics, a rich variety of nuclear structures and reactions emerge from strong interactions between protons and neutrons ring1980nuclear . In cold atoms, strongly correlated few- and many-body systems can be realized by tuning the s -wave scattering length between the atoms via a Feshbach resonance Inouye et al. 1998 ; Chin et al. 2010 , enabling the experimental observation of the Efimov states Efimov 1970 ; Efimov 1973 ; Kraemer et al. 2006 ; Braaten and Hammer 2006 ; Greene et al. 2017 ; Naidon and Endo 2017 ; d2018few ; Endo et al. 2025 . The Efimov states are peculiar three-body bound states characterized by Borromean binding and discrete scale invariance, that manifest as a quantum anomaly in the underlying field theory Bedaque et al. 1999 ; Braaten and Hammer 2006 ; Greene et al. 2017 ; Naidon and Endo 2017 ; d2018few ; Endo et al. 2025 . To understand the physical nature of the Efimov states and to gain insight into many-body physics from a few-body perspective, few-body systems with large s -wave scattering lengths have been extensively studied both experimentally Kraemer et al. 2006 ; Zaccanti et al. 2009 ; Gross et al. 2009 ; Pollack et al. 2009 ; Ferlaino et al. 2009 ; Zenesini et al. 2013 ; Pires et al. 2014 ; Tung et al. 2014 and theoretically Esry et al. 1999 ; von Stecher et al. 2009 ; von Stecher 2010 ; Gattobigio and Kievsky 2014 ; Naidon et al. 2012 ; Hiyama and Kamimura 2014 . These studies have revealed that, in addition to three-body bound states, four-, five-, and six-body states tied to Efimov trimers also appear universally Pollack et al. 2009 ; Ferlaino et al. 2009 ; Zenesini et al. 2013 ; von Stecher et al. 2009 ; von Stecher 2010 ; Gattobigio and Kievsky 2014 ; Hiyama and Kamimura 2014 . This hierarchy of few-body clusters provides an important step toward a unified understanding of how few-body systems evolve into many-body systems as the particle number increases Grebenev et al. 1998 ; Wenz_2013 ; von Stecher 2010 ; Gattobigio and Kievsky 2014 ; Yan and Blume 2015 ; Levinsen et al. 2017 . As Efimov binding is a highly quantum-mechanical phenomenon that defies classical description, it is an ideal testbed for benchmarking novel numerical methods. Indeed, strongly correlated few-body problems have been accurately solved with high precision using the correlated Gaussian basis Blume and Daily 2010 ; Daily and Blume 2012 ; von Stecher et al. 2009 , adiabatic hyper-spherical expansion von Stecher et al. 2009 , Gaussian expansion Naidon et al. 2012 ; Hiyama and Kamimura 2014 , and hyper-spherical harmonics expansion Gattobigio and Kievsky 2014 . As the Efimov states appear not only in cold atoms but also universally in a wide range of systems with large s -wave scattering lengths, such as weakly bound nuclear systems Hammer and Platter 2010 ; Endo et al. 2025 ; Tanihata et al. 2008 ; Hove et al. 2018 , 4 He clusters Kunitski , and magnetic systems Nishida et al. 2013 , establishing a neutral-network method for Efimov physics would lay the foundation toward a universal computational framework for exploring strongly correlated few- and many-body problems across various physical fields and scales. Here, we develop a neural-network method for computing not only the ground state but also excited states in strongly interacting few-body systems in uniform space and demonstrate that it accurately captures the Efimov states and their associated few-body clusters. Our implementation employs a fully connected feedforward neural network BHM ; Cai with suitably chosen Jacobi coordinates as inputs. The ground state is obtained by minimizing the variational energy with respect to the network parameters, and the first excited state is computed using the orthogonal projection method Choo . For three- to six-body systems of identical bosons at the unitary limit, as well as for mass-imbalanced three-body fermionic systems composed of two identical fermions and another particle, the ground and first excited states are robustly obtained. The results are in excellent agreement with previous results Gattobigio and Kievsky 2014 ; Blume and Daily 2010 ; Daily and Blume 2012 , achieving comparable or better accuracy. For three-body systems, we reproduce the expected universal features of the Efimov states, such as the discrete scale invariance, characteristic wave-function structures, and critical behavior as the mass ratio approaches 13.606\ldots for fermions Efimov 1973 ; Petrov 2003 . Our work thus establishes the neural-network method as a highly accurate and versatile method for exploring strongly interacting few-body problems. The rest of this paper is organized as follows. Section II explains the method of neural-network quantum states. Sections III and IV study systems of identical bosons and two identical fermions with one particle, respectively. In each of these sections, the method and numerical results are presented. Section V gives the conclusions and suggestions for future research. II Neural-Network Quantum States To represent a wave function, we use a fully-connected feedforward neural network, which consists of an input layer, hidden layers, and an output layer Goodfellow . The input layer receives a real-valued vector, \bm{u}^{\mathrm{in}}=\left(u^{\mathrm{in}}_{1},u^{\mathrm{in}}_{2},\dots,u^{\mathrm{in}}_{N_{\mathrm{in}}}\right), (1) constructed from the particle coordinates, where N_{\mathrm{in}} denotes the number of input values. We will specify later how to construct \bm{u}^{\mathrm{in}} from the particle coordinates in a uniform space. The neural network contains L_{h} hidden layers. The units in the \ell -th hidden layer ( \ell=1,2,\dots,L_{h} ) are defined recursively as u^{(\ell)}_{i}=\sum_{j=1}^{N_{\ell-1}}W^{(\ell)}_{ji}\,f\left(u^{(\ell-1)}_{j}\right)+b^{(\ell)}_{i}, (2) where N_{\ell} denotes the number of units in the \ell -th layer, W^{(\ell)} is a real N_{\ell-1}\times N_{\ell} matrix, and b^{(\ell)} is a real N_{\ell} component vector. The input layer corresponds to \ell=0 . As an activation function f(x) , we adopt the SiLU (Sigmoid Linear Unit) function: f(x)=\frac{x}{1+e^{-x}}. (3) We found that this activation function allows more stable and accurate calculations than those obtained using other activation functions, such as the ReLU (Rectified Linear Unit) and sigmoid functions. The output layer gives the following two real values: u^{\mathrm{out}}_{k}=\sum_{i=1}^{N_{L_{h}}}W^{(L_{h})}_{ik}f\left(u^{(L_{h})}_{i}\right)\qquad(k=1,2). (4) In this paper, we use a neural network with L_{h}=3 and N_{\ell}=32 . Finally, the network outputs a single value A : A=u^{\mathrm{out}}_{1}\exp\left(u^{\mathrm{out}}_{2}\right), (5) so that A can take either sign with exponential nonlinearity. The input values \bm{u}_{\rm in} are made from the particle coordinates X . The many-body trial wave function \Psi is constructed using the network output A(X,W) in combination with X , where W denotes the network parameters. The wave function \Psi therefore depends on both X and W . According to the variational principle, the expectation value E(W) of the Hamiltonian \hat{H} for the trial wave function \Psi is not less than the true ground-state energy E_{0} : E(W)=\frac{\int\Psi^{*}\hat{H}\Psi dX}{\int|\Psi|^{2}dX}\geq E_{0}. (6) We minimize E(W) with respect to W using the variational Monte Carlo method. In the following sections, we rewrite E(W) in a form generally given by E(W)=\frac{\int F(P,W)dP}{\int|\Psi(P,W)|^{2}dP}, (7) where F is a function constructed from the network output A and the Jacobi coordinates P of the particle positions. By taking the Metropolis-Hastings sampling of the coordinates P with a probability distribution p(P,W)=\frac{|\Psi(P,W)|^{2}}{\int|\Psi(P,W)|^{2}dP}, (8) we can evaluate the multidimensional integration in Eq. ( 7 ) as \displaystyle E(W) \displaystyle= \displaystyle\int p(P,W)\frac{F(P,W)}{|\Psi(P,W)|^{2}}dP (9) \displaystyle\simeq \displaystyle\frac{1}{N_{s}}\sum_{n=1}^{N_{s}}\frac{F(P_{n},W)}{|\Psi(P_{n},W)|^{2}}, where N_{s} is the number of samples. Similarly, the gradient of the energy with respect to the network parameters is obtained as \frac{\partial E}{\partial W}\simeq\frac{1}{N_{s}}\sum_{n=1}^{N_{s}}\frac{1}{|\Psi|^{2}}\left(\frac{\partial F}{\partial W}-E\frac{\partial|\Psi|^{2}}{\partial W}\right). (10) Using this gradient, we update the network parameters using the Adam optimizer Adam to minimize the energy. The learning rate in the Adam optimizer is typically chosen to be 10^{-3} – 10^{-4} . We use the projection method to obtain the excited state. Namely, we define the following wave function Choo : \Psi(P,W)=\Psi_{1}(P,W)-\lambda(W)\Psi_{0}(P,W_{0}), (11) where \Psi_{0}(P,W_{0}) is the ground-state wave function obtained using the above method and \Psi_{1}(P,W) is a wave function constructed from a network that is different from that used for the ground state. The parameter \lambda in Eq. ( 11 ) is given by \lambda(W)=\frac{\int\Psi_{0}^{*}(P,W_{0})\Psi_{1}(P,W)dP}{\int|\Psi_{0}(P,W_{0})|^{2}dP}, (12) which is also evaluated using Monte Carlo sampling. The wave function \Psi in Eq. ( 11 ) is orthogonal to the ground state \Psi_{0} . By minimizing the energy with respect to the network parameters W , we obtain the first excited state. The gradient of the energy can be numerically calculated using the automatic differentiation provided by a machine-learning framework. III Bosonic Systems III.1 Method We first apply our neural-network method to systems of identical N bosons with mass m in three-dimensional free space. The position of the i -th particle is denoted by the Cartesian coordinate \bm{r}_{i}\in\mathbb{R}^{3} . The i -th and j -th particles interact with each other through the two-body potential V(r_{ij}) , where r_{ij}=|\bm{r}_{i}-\bm{r}_{j}| . The Hamiltonian for the system is given by \hat{H}=-\frac{\hbar^{2}}{2m}\sum_{i=1}^{N}\frac{\partial^{2}}{\partial\bm{r}_{i}^{2}}+\sum_{i13.606\ldots Efimov 1973 ; Petrov 2003 ; Blume and Daily 2010 ; Daily and Blume 2012 . Such highly mass-imbalanced fermionic systems have recently been realized in cold atom mixtures of Er-Li Schäfer et al. 2023 and Dy-Li Xie et al. 2025 , where the fermionic Efimov states are expected to appear in the \ell^{\Pi}=1^{-} angular-momentum and parity sector Oi et al. 2024 ; Oi and Endo 2025 . IV.1 Method We consider a three-body system composed of two identical fermions located at \bm{r}_{1} and \bm{r}_{2} with mass M , and another particle located at \bm{r}_{3} with mass m . The two-body interaction is introduced only between each fermion and the third particle; that is, no interaction is assumed between the two identical fermions, since there is no s -wave interaction between them. The Hamiltonian for this system is given by \displaystyle\hat{H}_{F} \displaystyle= \displaystyle-\frac{\hbar^{2}}{2M}\left(\frac{\partial^{2}}{\partial\bm{r}_{1}^{2}}+\frac{\partial^{2}}{\partial\bm{r}_{2}^{2}}\right)-\frac{\hbar^{2}}{2m}\frac{\partial^{2}}{\partial\bm{r}_{3}^{2}} (34) \displaystyle+V\left(r_{13}\right)+V\left(r_{23}\right). Using the Jacobi coordinates, \displaystyle\bm{\rho}_{1} \displaystyle= \displaystyle\sqrt{\frac{M+m}{2m}}\left(\bm{r}_{2}-\bm{r}_{1}\right), (35a) \displaystyle\bm{\rho}_{2} \displaystyle= \displaystyle\sqrt{\frac{2(M+m)}{2M+m}}\left(\bm{r}_{3}-\frac{\bm{r}_{1}+\bm{r}_{2}}{2}\right), (35b) we can eliminate the center-of-mass motion from the Hamiltonian and obtain \hat{H}_{F}^{\prime}=-\frac{\hbar^{2}}{2\mu}\frac{\partial^{2}}{\partial\bm{\rho}_{1}^{2}}-\frac{\hbar^{2}}{2\mu}\frac{\partial^{2}}{\partial\bm{\rho}_{2}^{2}}+V_{\mathrm{PT}}\left(r_{13}\right)+V_{\mathrm{PT}}\left(r_{23}\right), (36) where \mu=Mm/(M+m) is the reduced mass, r_{13}=|\sqrt{m}\bm{\rho}_{1}+\sqrt{2M+m}\bm{\rho}_{2}|/\sqrt{2(M+m)} , and r_{23}=|\sqrt{m}\bm{\rho}_{1}-\sqrt{2M+m}\bm{\rho}_{2}|/\sqrt{2(M+m)} . The Hamiltonian in Eq. ( 36 ) commutes with the angular-momentum operator, \hat{L}=\sum_{j=1}^{2}\left(\bm{\rho}_{j}\times\frac{\hbar}{i}\frac{\partial}{\partial\bm{\rho}_{j}}\right). (37) The eigenstates of the Hamiltonian \hat{H}_{F}^{\prime} can therefore be characterized by the angular-momentum quantum numbers. We express the Jacobi vectors \bm{\rho}_{1} and \bm{\rho}_{2} by their lengths \rho_{1} and \rho_{2} , the angle \theta between them, and their overall rotation by the Euler angles \phi , \theta^{\prime} , and \phi^{\prime} as \bm{\rho}_{1}=R_{z}(\phi^{\prime})R_{y}(\theta^{\prime})R_{z}(\phi)\left(\begin{array}[]{c}0\\ 0\\ \rho_{1}\end{array}\right)=\rho_{1}\left(\begin{array}[]{c}\sin\theta^{\prime}\cos\phi^{\prime}\\ \sin\theta^{\prime}\sin\phi^{\prime}\\ \cos\theta^{\prime}\end{array}\right), (38) \bm{\rho}_{2}=R_{z}(\phi^{\prime})R_{y}(\theta^{\prime})R_{z}(\phi)\left(\begin{array}[]{c}\rho_{2}\sin\theta\\ 0\\ \rho_{2}\cos\theta\end{array}\right)=\rho_{2}\left(\begin{array}[]{c}\cos\theta\sin\theta^{\prime}\cos\phi^{\prime}+\sin\theta(\cos\phi\cos\theta^{\prime}\cos\phi^{\prime}-\sin\phi\sin\phi^{\prime})\\ \cos\theta\sin\theta^{\prime}\sin\phi^{\prime}+\sin\theta(\cos\phi\cos\theta^{\prime}\sin\phi^{\prime}+\sin\phi\cos\phi^{\prime})\\ \cos\theta\cos\theta^{\prime}-\sin\theta\cos\phi\sin\theta^{\prime}\end{array}\right), (39) where R_{x,y,z}(\alpha) is the rotation matrix, with rotation by an angle \alpha around the x , y , or z axis. In this coordinate system, the operator \hat{\bm{L}}^{2} can be represented only by the Euler angles (it does not include \rho_{1} , \rho_{2} , and \theta ) Breit 1930 . We denote the eigenvalues of \hat{\bm{L}}^{2} and \hat{L}_{z}=\hbar\partial/(i\partial\phi^{\prime}) as \hbar^{2}\ell(\ell+1) and \hbar m_{z} , respectively, and the corresponding eigenfunction as Y_{n,\ell,m_{z}}(\phi,\theta^{\prime},\phi^{\prime}) , where the index n identifies the eigenstates within each angular-momentum subspace. The energy is degenerate with respect to m_{z} due to the rotational symmetry and we only consider the m_{z}=0 states. We numerically confirmed that the \ell=1 subspace with odd parity has the lowest energy, consistent with the zero-range results, and therefore focus on this channel throughout this section. We note, however, that our formalism can be systematically extended to other angular-momentum and parity channels. The parity transformation, \bm{\rho}_{1}\rightarrow-\bm{\rho}_{1} and \bm{\rho}_{2}\rightarrow-\bm{\rho}_{2} , is rewritten as \theta^{\prime}\rightarrow\pi-\theta^{\prime} , \phi\rightarrow\pi-\phi , and \phi^{\prime}\rightarrow\phi^{\prime}+\pi . There are nine eigenfunctions with \ell=1 Breit 1930 . Among them, the eigenfunctions with m_{z}=0 and odd parity are given by Y_{1,1,0}=\cos\theta^{\prime},\qquad Y_{2,1,0}=\sin\theta^{\prime}\cos\phi. (40) The general form of the wave function can therefore be written as \displaystyle\psi(\rho_{1},\rho_{2},\theta;\theta^{\prime},\phi)={} \displaystyle f_{1}(\rho_{1},\rho_{2},\theta)\cos\theta^{\prime} (41) \displaystyle+f_{2}(\rho_{1},\rho_{2},\theta)\sin\theta^{\prime}\cos\phi, where f_{1} and f_{2} are unknown functions. We express the functions f_{1} and f_{2} by two different neural networks as f_{1}=A(P,W_{1}) and f_{2}=A(P,W_{2}) , respectively, where A is the network output given in Eq. ( 5 ). The wave function must be antisymmetric with respect to the exchange of the two identical fermions. In the Jacobi coordinates, the exchange of the fermions is expressed as \bm{\rho}_{1}\rightarrow-\bm{\rho}_{1} and \bm{\rho}_{2}\rightarrow\bm{\rho}_{2} , which are rewritten as \theta\rightarrow\pi-\theta , \theta^{\prime}\rightarrow\pi-\theta^{\prime} , and \phi\rightarrow-\phi , as found from Eqs. ( 38 ) and ( 39 ). Thus, in a manner similar to Eqs. ( 21 ) and ( 22 ), we construct the trial wave function as \Psi=\varphi A_{\rm asym}, (42) where \varphi=\phi(r_{13})\phi(r_{23}) (43) is the product of two-body correlations and A_{\rm asym}=\psi(\rho_{1},\rho_{2},\theta,\theta^{\prime},\phi)-\psi(\rho_{1},\rho_{2},\pi-\theta,\pi-\theta^{\prime},-\phi) (44) ensures antisymmetrization for the fermions. As in the bosonic case in Eq. ( 24 ), we can define the effective potential V_{\rm eff} arising from the derivative of \varphi . The energy integral is written as \int\Psi\hat{H}_{F}^{\prime}\Psi dP=\int dP\left[\frac{\hbar^{2}}{2\mu}\sum_{j=1}^{2}\varphi^{2}\left(\frac{\partial A_{\rm asym}}{\partial\bm{\rho}_{j}}\right)^{2}+V_{\rm eff}\Psi^{2}\right], (45) where dP=d\bm{\rho}_{1}d\bm{\rho}_{2} . The procedures used to optimize the network parameters are the same as those for the bosonic case. The first excited state is searched within the same subspace as the ground state, namely the \ell=1 and odd parity subspace. In the following numerical calculations, we use the Pöschl-Teller potential for the two-body interaction potential as V_{\mathrm{PT}}(r)=-\frac{\hbar^{2}}{\mu b^{2}}\frac{1}{\cosh^{2}\frac{r}{b}}, (46) whose s -wave scattering length is infinite, corresponding to the unitary limit. We use the zero-energy analytical solution of this potential in Eq. ( 30 ) as the two-body correlation \phi in Eq. ( 43 ). The two-body potential term in the Hamiltonian vanishes, as in the bosonic case, and the effective potential simplifies to V_{\rm eff}=-\frac{\hbar^{2}}{m}\bm{b}_{13}\cdot\bm{b}_{23}, (47) where \bm{b}_{ij} is defined in Eq. ( 33 ). IV.2 Results First, we focus on the mass ratio M/m=27.752 corresponding to the {}^{167}\mathrm{Er} - {}^{6}\mathrm{Li} system. A mixture of these ultracold atomic gases has been realized experimentally, in which interspecies Feshbach resonances were observed Schäfer et al. 2023 . By fine-tuning the s -wave scattering length to be large and achieving the unitary limit, the Efimov states involving identical fermions are expected to appear in the \ell=1 channel when the mass ratio is larger than the critical value, M/m\gtrsim 13.606 . While a strong dipole-dipole interaction between the Er atoms may affect the physical properties of the Efimov states of Er-Er-Li, we neglect this interaction in this work and only consider the short-range interactions between Er and Li, tuned to the unitary limit, and demonstrate that our neural-network method is useful for investigating fermionic Efimov physics. Figure 4: Optimization process of neural networks for system of two identical fermions and another particle with mass ratio M/m=27.752 . (a) Ground-state energy and (b) first excited-state energy, normalized by \hbar^{2}/(mb^{2}) . 10^{5} Monte Carlo samples are taken in each update. Auxiliary harmonic potential in Eq. ( 28 ) is applied during the first 1000 steps, with \omega=10^{-2}\hbar/(mb^{2}) for the ground state and \omega=10^{-3}\hbar/(mb^{2}) for the excited state. The optimization well converges for \sim 10^{5} steps in (a), while \sim 10^{6} steps are needed for (b). Figure 4 shows the convergence of the energies with respect to the network parameters’ update steps. For both the ground and first excited states, the energies converge to negative values, indicating the formation of bound states. The excited state exhibits larger statistical fluctuations than those for the ground state, reflecting its extended spatial structure, which makes the Monte Carlo integration more difficult, as in the bosonic case. After the convergence of the optimization, we estimate the energies with 10^{8}\times 10^{3} Monte Carlo samples, in a manner similar to the bosonic case, giving E_{0}=-1.880\times 10^{-2}\hbar^{2}/(mb^{2}) for the ground state and E_{1}=-4.37\times 10^{-4}\hbar^{2}/(mb^{2}) for the first excited state. The ratio \sqrt{|E_{0}/E_{1}|}\simeq 6.6 is in reasonable agreement with the universal discrete scale factor predicted from the zero-range theory, \sqrt{|E_{0}/E_{1}|}\simeq 7.19 Efimov 1973 ; Braaten and Hammer 2006 ; Greene et al. 2017 ; Naidon and Endo 2017 ; d2018few ; Petrov 2003 for M/m=27.752 . The small deviation is attributed to the finite-range effects of the potential and the limited validity of the low-energy condition, which often deteriorate the universal description of the ground Efimov trimer. Figure 5: Probability distribution P(R) as function of hyperradius R for system of two identical fermions and one particle with mass ratio M/m=27.752 : (a) ground state and (b) first excited state. The insets show the probability distributions for the position of the third particle when the two identical fermions are fixed at the positions indicated by the crosses. Figures 5 (a) and 5 (b) show the probability distributions P(R) with respect to the hyperradius R for the ground and first excited states, respectively. As in the bosonic case, the two distributions have similar shapes with different spatial scales, exhibiting the discrete scale invariance in Efimov physics. The scale ratio is \simeq 5 , which is consistent with the zero-range prediction of \simeq 7.19 . In the first excited state [Fig. 5 (b)], a node appears in the short-range region, makeing the state orthogonal to the ground state. The insets in Fig. 5 show the probability distribution for the position of the third particle when the two identical fermions are fixed. The results show the same qualitative features as those for the bosonic three-body systems in Fig. 3 , demonstrating that the neural network can capture the spatial structure of fermionic Efimov states. Figure 6: Ground-state energy [normalized by \hbar^{2}/(mb^{2}) ] of system of two identical fermions and one particle as function of mass ratio M/m . Circles show the results obtained using the neural-network method for the Pöschl-Teller potential at the unitary limit in Eq. ( 46 ). Triangles show the results obtained using the neural-network method for the Gaussian potential in Eq. ( 48 ) at the unitary limit. Squares show the results obtained using the explicitly-correlated Gaussian-basis method for the same Gaussian potential Blume and Daily 2010 ; Daily and Blume 2012 . The vertical dotted line indicates M/m=13.606 . We also perform the calculation for variable mass ratios and obtain the dependence of the ground-state energy on the mass ratio M/m , as shown by the circles in Fig. 6 . As expected from the zero-range Efimov theory Efimov 1973 ; Braaten and Hammer 2006 ; Greene et al. 2017 ; Naidon and Endo 2017 ; d2018few , the binding energy increases monotonically with M/m . The data extrapolate to the disappearance point of the bound state, consistent with the critical mass ratio of 13.606\ldots Efimov 1973 ; Petrov 2003 , confirming that the neural-network method can reliably describe fermionic few-body systems even near the binding threshold. We finally consider the case in which the two-body interaction potential is not a Pöschl-Teller potential. We consider the Gaussian potential V_{G}(r_{ij})=V_{0}e^{-r_{ij}^{2}/b^{2}}, (48) where b characterizes the interaction range. The value of V_{0} is taken to be -1.342\hbar^{2}/(\mu b^{2}) , corresponding to the depth at which the two-body ground bound state is about to appear and hence to the unitary limit. Although the analytical two-body solution is unavailable for the Gaussian potential, we can still perform the neural-network calculation by using the two-body correlation of the Pöschl-Teller potential, exploiting the similar, though not identical, short-range correlations of the two potentials. More specifically, in solving the problem with the Gaussian potential, we use the trial wave function in Eqs. ( 42 ) and ( 43 ) with the Pöschl-Teller two-body solution in Eq. ( 30 ). The effective potential changes from Eq. ( 47 ) to \displaystyle V_{\rm eff} \displaystyle= \displaystyle-\frac{\hbar^{2}}{m}\bm{b}_{13}\cdot\bm{b}_{23}+V_{G}(r_{13})+V_{G}(r_{23}) (49) \displaystyle-V_{\rm PT}(r_{13})-V_{\rm PT}(r_{23}). The obtained ground-state energies are shown by the triangles in Fig. 6 . They are in excellent agreement with those obtained using the explicitly-correlated Gaussian-basis method Blume and Daily 2010 ; Daily and Blume 2012 for the same two-body Gaussian potential (squares in Fig. 6 ). This agreement is attributed to the ability of the neural network to compensate for the relatively small differences between the short-range two-body correlations of the Pöschl-Teller and Gaussian potentials, thereby yielding accurate energies. The above results suggest that the neural-network method is not limited to analytically solvable two-body potentials, but is applicable to problems with a wide variety of interaction potentials. V Conclusions and Discussion We developed the method of neural-network quantum states for exploring Efimov physics in few-body quantum systems in three-dimensional continuous space. By using the two-body solution to efficiently incorporate the strong two-body correlations at the unitary limit into the variational wave function, the convergence and accuracy are significantly improved. For systems of identical N bosons, as well as the mass-imbalanced three-body fermionic system, the ground and first excited states are obtained with an accuracy comparable to or better than previously reported results. The ratio between the ground and first excited-state energies in the three-body systems agrees well with the values predicted by the universal Efimov theory. In addition, our method reproduces the characteristic shapes of the Efimov states’ wave functions and captures the disappearance of the Efimov state as the mass ratio approaches the critical value of 13.606\ldots for mass-imbalanced fermions. These results demonstrate that the neural-network quantum states provide a versatile and accurate approach for exploring few-body systems near the unitary limit. In most of the numerical calculations presented in this paper, we used the Pöschl-Teller potential, for which the analytical form of the two-body correlation \phi and its derivative \bm{b}_{ij} are available, with which the convergence is accelerated. However, the proposed method is not restricted to analytically solvable potentials. Indeed, leveraging the fact that the two-body correlations \phi exhibit similar, though not identical, short-range behavior, we used the two-body correlation of the Pöschl-Teller potential to study a three-body system with a Gaussian potential. The neural networks are found to compensate for the small differences in the two-body correlations of different potentials and to converge rather efficiently to the accurate values obtained using an established method Blume and Daily 2010 ; Daily and Blume 2012 . Alternatively, one may employ the numerically obtained \phi and \bm{b}_{ij} with appropriate interpolation. These schemes would allow solving few-body problems with general interaction potentials, including regimes away from the unitary limit considered here, or dipolar types of interactions whose long-range and anisotropic nature leads to exotic few-body clusters Wang et al. 2011a ; Wang et al. 2011b ; Oi et al. 2024 ; Oi and Endo 2025 ; Shi et al. 2026 ; Ohishi . Acknowledgements. We thank D. Blume for providing the data shown as squares in Fig. 6 , which were calculated using the same methods as used in Refs. Blume and Daily 2010 ; Daily and Blume 2012 . HS was supported by JSPS KAKENHI Grant Numbers JP23K03276 and JP26K00638. SE acknowledges support from Matsuo Foundation and JSPS KAKENHI Grant Number JP25K00217.