acceptodds
Under review as a conference paper at ICLR 2027

Symmetry-Aware Residual Learning of Few-Body Molecular Wave Functions

Abstract

Ultracold polar molecules are a highly controllable platform for engineering quantum matter, but modeling their few-body properties remains computationally challenging. Their anisotropic long-range interactions and rich internal structure lead to high-dimensional eigenvalue problems for which conventional grid-based solvers are largely restricted to two-molecule systems. To address this limitation, we introduce a physics-informed neural variational method that exploits known physical structure of the molecular wave function. Our key idea is to factor the wave function into analytic and neural components. The analytic component encodes rotational symmetry and asymptotic decay, while the neural residual learns the remaining intermediate-range behavior. This decomposition reduces the effective dimensionality of the problem without relying on ad hoc approximations or problem-specific discretization grids. We benchmark our method against high-accuracy two-molecule solvers and recover both ground and excited bound states with relative errors at the percent level. We also validate the numerical building blocks required for extension to three-body systems. Specifically, we test the angular-channel decomposition and two-coordinate radial representation in settings where reliable benchmarks are available. This provides a systematic route to larger molecular systems and opens opportunities to study correlations and dynamics in interacting quantum matter.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
ML for Physics, PINN, Neural wave-function

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FM5gOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a neural variational method for computing bound-state wave functions of ultracold polar molecules. It combines an analytical spherical-harmonic expansion with a WKB-inspired envelope, leaving a shared MLP to learn the remaining radial structure. The method is benchmarked against accurate two-body solvers for ground and excited states across interaction strengths, with ablations examining the physical priors and parametrization. A three-body formulation is also derived, although validation is limited to its numerical components rather than a complete three-body calculation.

Critical strengths

The representation is well motivated by the target physics: analytical angular decomposition and a WKB-inspired envelope reduce what the network must learn. This gives the design a clear connection to the problem, and the ablations support its benefits for optimization.

The two-body results provide credible validation against an accurate classical solver across interaction strengths, particularly for ground states. Testing excited states and multiple random seeds adds confidence that the method works beyond a single favorable example.

Critical weaknesses

The effect of angular-channel truncation is insufficiently validated. Both the neural method and the MDVR reference use l_max=8, so their agreement does not establish convergence with respect to the angular basis. Increasing this cutoff would help determine whether important physical contributions are omitted.

The WKB-inspired envelope uses a simplified potential, a proxy energy, and an empirical broadening factor. Although the ablations show optimization benefits, they do not sufficiently establish whether these choices affect final accuracy or wave-function tails. This limits confidence in the claim that the representation captures the correct asymptotic behavior.

Questions for authors

Could you provide convergence results with increasing l_max for both the neural method and the reference solver? Since both use lmax=8, their agreement does not establish that angular truncation errors are negligible.

Beyond the optimization trajectories in the envelope ablation, how sensitive are the final energies and wave-function tails to the proxy energy and broadening factor? Does the learned residual recover the expected asymptotic decay despite the deliberately broadened envelope?

These are requests to validate the current results, rather than suggestions for future work.

1 comment

We thank the Reviewer for their assessment of the physical motivation and two-body validation. We have added angular-cutoff and envelope-sensitivity studies to the revised Manuscript.

Angular-channel convergence

To test convergence in the angular basis, we performed the ground-state optimization with lmax⁡∈2,4,6,8,10,12l_{\max}\in\\{2,4,6,8,10,12\\} for ν∈50,100,500,1000 MHz\nu\in\\{50,100,500,1000\\}\thinspace\text{MHz}. We used three random seeds for each configuration, keeping all other settings fixed. The table below provides the seed-average relative error with respect to MDVR (computed at lmax⁡=8l_{\max}=8), and the change in the neural energy compared to the previous cutoff value, which measures convergence independently of the reference.

lmax⁡l_{\max}Relative error vs MDVR (%)Change from the previous lmax⁡l_{\max} (%)
240 – 43–
42.1 – 6.337 – 42
60.04 – 0.492.0 – 5.9
80.0003 – 0.0140.04 – 0.48
100.00005 – 0.019 (below)0.0004 – 0.033
120.0002 – 0.017 (below)0.0001 – 0.012

The energy converges quickly with lmax⁡l_{\max}. In particular, the error drops from about 40% for lmax⁡=2l_{\max}=2 to below 0.5% for lmax⁡=6l_{\max}=6. Beyond the cutoff used in the paper, raising lmax⁡l_{\max} from 8 to 10 changes the energy by at most 0.03%, and from 10 to 12 by at most 0.012% with no consistent sign, which is at the level of the seed-to-seed variation. The neural energy can therefore be considered converged (in the angular basis sense) at lmax⁡=8l_{\max}=8.

The same results also bound the truncation error of the reference. At lmax⁡=10l_{\max}=10 and 12 the neural energy lies at most 0.02% below the MDVR energy, so the channels omitted by MDVR at lmax⁡=8l_{\max}=8 contribute at most about 0.02% to the ground-state energy. Both truncation errors are therefore small compared to the differences discussed in the paper. The full results, including training dynamics for each lmax⁡l_{\max}, are in Appendix E.3.2.

We also tested angular convergence in the newly introduced three-body calculation. Increasing the dual-angular cutoff from lmax⁡,Lmax⁡=4,4\\{l_{\max},L_{\max}\\}=\\{4,4\\} to 6,6\\{6,6\\} lowers the ground-state energy by ≈15\approx 15\\%. The subsequent enlargement to 8,8\\{8,8\\} changes the energy by ≈3\approx 3\\%, whereas the final enlargement to 10,8\\{10,8\\}, or from 145 to 190 channels, changes it by only ≈0.25\approx 0.25\\%. This decreasing sensitivity supports the cutoff 10,8\\{10,8\\} used in the three-body benchmarks (both neural and the reference 2D DVR). The results are included in Appendix E.3.2.

Envelope sensitivity and learned tails

We expanded the envelope ablation to β∈0,0.001,0.01,0.1,1\beta\in\\{0,0.001,0.01,0.1,1\\}, where β=0\beta=0 removes the envelope decay and β=1\beta=1 retains the literal WKB envelope. We tested ground states for ν∈50,100,500,1000 MHz\nu\in\\{50,100,500,1000\\}\thinspace\text{MHz} and first excited states for ν∈400,500,700,1000 MHz\nu\in\\{400,500,700,1000\\}\thinspace\text{MHz}, with three random seeds per configuration. All runs use the same fixed logarithmic-scale grid, separating the effect of the envelope on wave function from its effect on sampling.

For our main choice (β=0.01)(\beta=0.01), the seed-averaged energy error relative to MDVR is at most 0.0005% for ground states and 0.004% for excited states. Removing the envelope decay also yields accurate final energies, but convergence is several times slower. In contrast, the literal envelope is too restrictive: ground-state errors reach about 2%, and the excited-state calculation remains above the dissociation threshold at the two lowest tested frequencies. Section 3 of the General Reply summarizes the full sensitivity results.

We also directly compared the learned tails with MDVR beyond the outer turning point, for ground states at ν=50 MHz\nu=50\thinspace\text{MHz} and ν=500 MHz\nu=500\thinspace\text{MHz} and excited states at ν=400 MHz\nu=400\thinspace\text{MHz} and ν=1000 MHz\nu=1000\thinspace\text{MHz}. For β=0.01\beta=0.01, the learned wave function falls well below the broadened envelope and follows the reference MDVR decay over several orders of magnitude. The residual therefore supplies state-dependent decay that is absent from the envelope itself. At very small wave function amplitudes, the curves depart and numerical oscillations appear. Nevertheless, the evidence supports recovery of the tail over the resolved range.

These tests retain the prescribed proxy energy EWKB=min⁡rV(r)/2E_{\text{WKB}}=\min_r V(r)/2, and do not vary it independently. They show that the residual compensates for its decay-rate mismatch at the chosen broadening, including for a weakly bound excited state. Appendix E.3.1 presents the energy results, training curves, and direct tail comparisons.

Reviewer_4BRZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I understand this paper as a structured neural variational solver for bound states of microwave-shielded polar molecules. It combines analytical angular decomposition, a tempered WKB-inspired envelope, shared neural radial functions, and a fixed sampling pool. The experiments compare two-body ground and first excited states with MDVR, investigate component ablations and optimization behavior, and report repeated runs at representative field strengths. A three-body formulation is presented, but its numerical validation consists of separate component tests on two-body problems.

Critical strengths

Physically motivated representation. The spherical-harmonic expansion and WKB-inspired envelope provide angular structure and radial priors, allowing the shared network to focus on the remaining radial functions while retaining channel coupling. This division between analytical and learned components is sensible.

Useful two-body validation. The field-strength sweep, wave-function comparisons, and five-seed results at representative settings provide evidence beyond isolated examples. The independent Weinberg-expansion check strengthens confidence in the ground-state MDVR references.

Critical weaknesses
  1. No comparison with closely related neural solvers. Li et al. already combine angular/multipole expansions with neural channel functions. Without an adapted baseline on the same Hamiltonian, I cannot determine the additional value of the proposed design. Reference: Li et al., Solving two and three-body systems with deep neural networks, 2025.

  2. The efficiency advantage is not established. MDVR remains faster. The reported speedup over the cylindrical-coordinate network also changes integration, differentiation, and batch size, so it does not isolate the benefit of the proposed representation.

  3. The sampling contribution is insufficiently validated. Logarithmic sampling performs comparably to WKB sampling, and no comparison with sampling from the evolving wave function is provided. The evidence does not establish a distinct benefit from fixing the sampling pool.

  4. Higher-body applicability remains unverified. All completed calculations concern one two-body potential family; the two-coordinate experiment is also two-body. A formal three-body extension does not demonstrate that this construction scales better than other neural variational methods.

Questions for authors

On the same Hamiltonian, how does the proposed method compare with an adapted neural few-body baseline in accuracy, total runtime, and memory? Results at comparable computational budgets or target accuracy would help establish its practical advantage.

Can the authors provide a complete coupled three-body benchmark that tests the benefit of the proposed construction as particle number increases, including angular-channel convergence and computational cost?

With the trained networks held fixed, do the reported energies remain stable under independent, higher-accuracy numerical evaluation? Please explain the ground-state estimates below both reference solvers.

2 comments

We thank the Reviewer for their insightful comments. We appreciate that they found the representation physically well motivated and the two-body validation useful.

Critical Weaknesses:

Two of the critical weaknesses are also subjects of Reviewer's questions, hence they are discussed in the next section.

Efficiency compared to MDVR and 2D DVR

We agree that MDVR remains faster for the two-body problems tested, and that our results do not establish a computational advantage over MDVR. Our motivation concerns extending the representation to three-body states, where the same mapping strategy faces additional challenges.

For two bodies, MDVR uses 512 radial points over r∈[50,105]⋅a0r\in[50,10^5]\cdot a_0 with lmax⁡=8l_{\max}=8. Mapping concentrates points in the interaction region while coarsening the grid at the extended tail. By comparison, an illustrative uniform-grid DVR with spacing 10 a010\,a_0 would require approximately 10410^4 radial points. Including five even-ll partial waves gives a basis size of approximately 50,000 and requires 20 GB to store a fully dense, real double-precision Hamiltonian, excluding diagonalization workspace.

For three bodies, applying the same domain and spacing to both Jacobi radii produces 10810^8 radial grid pairs. With our refined 190-channel angular basis, the total basis size reaches approximately 1.9×10101.9\times10^{10}. A single real double-precision wavefunction vector would require approximately 152 GB, while the fully dense Hamiltonian would require approximately 2.89×10212.89\times10^{21} bytes, making dense storage impractical. These are illustrative storage estimates for 2D DVR, not convergence benchmarks.

Additionally, in the three-body case, radial mapping is more difficult than for two bodies, because the pair separations, ∣R±r/2∣|\mathbf R\pm\mathbf r/2|, can remain small even when both Jacobi radii are large. Coarsening the grids at large distances can therefore miss short-range pair correlations.

This motivates representations that avoid explicit storage of the full radial tensor-product grid. Although specialized classical methods can also reduce this storage burden, our approach offers flexibility of a continuous neural representation when extending to higher-body systems. Additional Jacobi coordinates become network inputs, while network size and quadrature resolution provide separately adjustable controls over representation and integration costs. This avoids explicitly constructing the full radial tensor-product basis or designing increasingly complicated coordinate mappings. The practical advantage is therefore a straightforward route to higher-body calculations with controllable memory requirements, supported here by the complete three-body demonstration.

Sampling strategy contribution

The original Appendix already compared NN performance on fixed grids. We extended it with the requested sampling from the evolving wave function (MCMC redraws from ∣Ψw∣2|\Psi_{w}|^2 every 5,000 or 500 steps) and a broader parameter coverage across four frequencies, three seeds and several grid sizes. All grids have 5⋅1045\cdot 10^4 points, and energies are evaluated on an independent dense grid. Ranges are over ν∈50,100,500,1000\nu\in\\{50,100,500,1000\\} MHz.

GridError vs MDVR (%)Steps to 1% convergence (of 15,000)
Uniform (fixed)0.5 – 6.23,300 – 12,400
Log-uniform (fixed)0.001 – 0.003590 – 1,500
WKB (fixed)0.001 – 0.010760 – 2,100
Model, every 5,000 steps0.2 – 7.910,300 – 13,300
Model, every 500 steps0.6 – 2613,400 – 15,000

Fixing the grid clearly helps. Resampling is orders of magnitude less accurate and barely converges, because each redraw changes the discrete objective and can push the energy into the continuum.

We agree that log-uniform sampling performs comparably here, since this radial problem happens to be close to logarithmic in rr. That transformation has to be found by hand for each problem, whereas the WKB-inspired density follows directly from the potential, which matters in multi-coordinate settings. Full results, including training curves and grid-size dependence, are in Appendix E.3.3.

Questions:

Comparison to related neural solvers

Li et al. (2025) demonstrate neural radial-channel representations for nuclear bound states in two- and three-body calculations. Our contribution is to introduce and validate numerical choices motivated by a different interaction potential. Short-range shielding and anisotropic long-range dipolar attraction produce states whose radial extent changes substantially with field strength. These features impose demands on both the wave-function representation and numerical integration that Li et al.’s nuclear benchmarks do not validate for the interaction considered here.

We incorporate the potential into both components: a tempered WKB envelope supplies a field-dependent radial profile for the neural residual to refine, while a potential-derived quadrature grid distributes resolution across the interaction region and extended tail. Our ablations assess these contributions separately on the same Hamiltonian. At equal point count, replacing uniform spacing with WKB spacing reduces dimer ground-state energy errors relative to MDVR from 0.5–6.2% to 0.001–0.010% across four field strengths. Logarithmic spacing gives comparable accuracy (0.001–0.003%), indicating that resolving the different radial scales accounts for the improvement in this test. The WKB construction provides a prescription for choosing point locations directly from the potential. With the logarithmic grid held fixed, training with and without envelope decay yields accurate ground- and excited-state energies, but the tempered envelope reaches that accuracy in fewer optimization steps.

We validate the resulting formulation against MDVR for ground and excited dimer (N=2)(N=2) states and compare trimer (N=3)(N=3) ground states with an independent 2D DVR calculation. Together, these benchmarks and ablations establish applicability to the target interaction and quantify the benefits of the numerical design. Section 2 of the General Reply provides the detailed comparison and supporting results.

Higher-body applicability and three-body benchmark

We have now performed a complete field-linked three-body calculation, testing the angular decomposition and two-coordinate radial representation together. The Hamiltonian includes all three pair interactions. Two pair separations, ∣R±r/2∣|\mathbf R\pm\mathbf r/2|, depend on both Jacobi vectors, so their potentials couple the angular channels and introduce correlations between the two radii. Within each channel, the network learns a function of both radii jointly.

We compared neural trimer ground-state energies with an independent 2D DVR calculation using the same Hamiltonian and angular truncation. The energies agree within 1% at nine of the ten tested field strengths, with discrepancies of 0.094–0.785% at the tested frequencies from 97.45 to 1431.16 MHz. The largest discrepancy, 2.48%, occurs at the weakest field tested, 68.31 MHz. Separately, increasing the angular cutoff from lmax⁡,Lmax⁡=8,8\\{l_{\max},L_{\max}\\}=\\{8,8\\} to 10,8\\{10,8\\} enlarges the basis from 145 to 190 channels and changes the neural ground-state energy by at most 0.25% at the frequencies tested for this refinement.

These results demonstrate that the formulation handles jointly dependent radial functions and coupled angular channels in a complete three-body Hamiltonian. Section 1 of the General Reply provides the full energy comparison.

Independent evaluation of trained networks

During training, the energy is estimated by quadrature on the training points, for which the network is optimized. We held the trained ground-state networks fixed (ν∈50,100,500,1000\nu\in\\{50,100,500,1000\\} MHz, three seeds) and re-evaluated their energy on independent logarithmic grids of 10410^4, 10510^5 and 10610^6 points, a uniform grid of 10610^6 points, and with exact derivatives by automatic differentiation. The table gives the seed-averaged relative error with respect to MDVR in %.

Evaluationν=50\nu=50 MHzν=100\nu=100 MHzν=500\nu=500 MHzν=1000\nu=1000 MHz
Training grid0.000510.000260.000330.00039
All independent grids0.000390.000270.00360.014

All independent evaluations agree to within 5×10−55\times10^{-5}% and lie above MDVR, as the variational principle requires. The training-grid estimate is up to 0.023% lower and hides the growth of the error with ν\nu. Energies below the reference therefore came from evaluating on the training points, where the network exploits quadrature errors, together with mini-batch fluctuations and an envelope evaluated on those same points. Details are placed in Appendix E.3.4.

Reviewer_RBjcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Ultracold polar molecules are a highly controllable platform for engineering quantum matter, but modeling their few-body properties remains computationally challenging because of their anisotropic long-range interactions. This paper introduces a physics-informed neural variational method that exploits known physical structure by factoring the molecular wave function into analytic components and a learned neural residual. Translational invariance is handled using Jacobi coordinates, the angular dependence is represented using spherical harmonics, and a tempered WKB-inspired envelope incorporates decay behavior. A shared MLP represents the remaining radial structure across angular channels. Ground states are obtained by minimizing the variational energy, while excited states are computed through orthogonalization against previously obtained states.

The contributions center on this structured residual representation, a physics-informed initialization and fixed-grid sampling procedure, and formulations for both two- and three-body systems. The method is benchmarked on two-body field-linked bound states across a broad range of Rabi couplings, with reported mean relative energy errors of 0.3% for ground states and 11.4% for excited states. Ablations examine the effects of the envelope, neural parametrization, and analytical angular decomposition. The three-body formulation is supported by separate two-body tests of its numerical building blocks; no actual three-body calculation is presented.

Critical strengths
  • Physically motivated representation for a meaningful application. The combination of analytical angular treatment, a decay envelope, and a flexible neural residual directly addresses the coexistence of anisotropy, localization, and extended wave-function tails. The ablations provide evidence that incorporating this structure improves optimization, particularly in shallow binding regimes. This is the strongest positive aspect of the work: the representation follows identifiable properties of the target problem, and its usefulness is supported by numerical results.

  • Convincing evidence of two-body ground-state feasibility. The comparison with MDVR covers a continuous range of interaction strengths, including both spatially extended and tightly localized states. The ground-state energy agreement demonstrates that the representation can accommodate substantially different length scales within the chosen potential family. This establishes a credible application-specific solver, although the excited-state accuracy and higher-body applicability are less well established.

Critical weaknesses
  • The experiments do not establish the central computational advance motivating the paper. The introduction identifies the difficulty of going beyond two-molecule calculations as the main bottleneck. However, every complete calculation remains a two-body problem for which an established solver already provides accurate results and is acknowledged to be faster. Section 6.3 tests angular decomposition in one two-body setting and a two-coordinate representation in another representation of the same two-body problem. These checks do not establish that the combined framework can handle the inter-coordinate correlations, coupled angular channels, and sampling difficulties of an actual three-body Hamiltonian. Consequently, the paper demonstrates an alternative two-body implementation without establishing an expanded computational capability or a new physical result. This gap is consequential because higher-body applicability supplies much of the proposed method’s significance.

  • The methodological novelty over closely related neural variational solvers is insufficiently established. Partial-wave decomposition followed by neural parametrization and variational optimization of coupled channels already appears in work on the deuteron, including Keeble and Rios and Rozalen Sarmiento et al. Li et al., which this paper cites, also treat tensor-coupled interactions and perform actual three-body calculations within a truncated channel space. The clearest additions here are therefore the particular WKB-inspired envelope and its associated sampling procedure. Their practical usefulness is plausible, but the current evidence does not establish a substantial methodological advance. In particular, the matrix-versus-MLP ablation changes spatial smoothness, sharing across radial positions, and sharing across angular channels simultaneously, so it does not isolate the claimed benefit of cross-channel sharing. The symmetry interpretation also needs correction. With the external-field direction fixed, the potential in Eq. (5) is not invariant under general SO(3) rotations. Spherical harmonics remain a valid and useful angular basis, but their use should be distinguished from enforcing full rotational symmetry of this Hamiltonian.

Questions for authors
  • Section 5.4 constructs training points from a fixed distribution q(r)∝∣ψWKB(r)∣2q(r)\propto|\psi_{\mathrm{WKB}}(r)|^2. What discrete Rayleigh quotient is actually optimized? Specifically, how are the energy, normalization, and orthogonalization integrals evaluated using these points: through importance weights proportional to (1/q(r)), or through equivalent nonuniform-grid quadrature weights?

  • The envelope uses a proxy energy E_wkb and a fixed tempering factor, whereas the physical asymptotic decay rate depends on each state’s actual eigenenergy. How does the chosen residual parametrization compensate for this mismatch, particularly for weakly bound excited states near threshold? Does the existing validation establish the accuracy of their long-range tails, or only their energies?

2 comments

We thank the Reviewer for their perspective on our work. We’re glad the physics-based design came through clearly and that they found the two-body benchmarks a credible validation across interaction strengths. Here, we address the highlighted weaknesses and answer Reviewer's questions.

Critical Weaknesses

Three-body (N=3)(N=3) calculation

We agree that our earlier tests validated the angular decomposition and the two-coordinate radial representation separately, leaving open whether they work together in a three-body Hamiltonian. We now report a complete three-body calculation using the same field-linked interaction as in the two-body case. The corresponding Hamiltonian includes all three pair interactions and the resulting coupling between dual-harmonic angular channels. Within each channel, the network learns a radial function that depends jointly on both Jacobi radii, allowing it to represent correlations between them.

We benchmarked the three-body ground-state energies against an independent 2D DVR calculation that uses the same angular decomposition. The two solvers agree to within 1% across a broad range of Rabi frequencies (ν∈[97.4525,1431.1639] MHz)(\nu\in[97.4525,1431.1639]\thinspace\text{MHz}), with the largest discrepancy (≈2.5%)(\approx2.5\%) at ν≈68 MHz\nu\approx68\thinspace\text{MHz}. Section 1 of the General Reply gives the full comparison and discusses the radial convergence of the DVR reference.

Comparison to related works

Our contribution is to introduce and validate numerical choices motivated by the interaction potential of our choice. The potential enters the variational objective directly and determines the spatial structure that the solver must resolve. Changing the potential therefore changes both the optimization problem and the requirements on representation and integration. Whereas Li et al. (2025) treat short-range nuclear interactions, our potential combines short-range shielding with anisotropic dipolar attraction decaying as r−3r^{-3}. Tuning the external field changes the potential well depth and the relevant radial scales, producing states ranging from weakly bound and spatially extended to tightly localized. The solver must resolve this variation while retaining angular-channel coupling and, for excited states, radial nodes and extended tails.

These demands motivate incorporating the potential into both the trial wave function and the quadrature. A tempered WKB-inspired envelope supplies a field-dependent radial profile, reducing the burden of learning strongly varying amplitudes while allowing the neural residual to recover state-dependent structure and decay. A fixed nonuniform quadrature grid uses the same potential information to distribute resolution across the interaction region and the extended tail. We test these choices through envelope and sampling ablations, benchmark ground and excited two-body states against MDVR, and compare complete three-body calculations with an independent 2D DVR reference. Li et al.’s nuclear benchmarks do not provide validation for the type of interaction considered here. Section 2 of the General Reply details the methodological differences and supporting results.

Rotational symmetries

We agree that the fixed external field breaks full rotational invariance of the Hamiltonian, thus we do not enforce this invariance. We use the phrase "rotational symmetries" in a different sense. In particular, spherical harmonics, which we use to treat angular integration and kinetic terms analytically, represent contributions from different rotational symmetry classes (labeled by index ll) to the expanded function. For instance, an Y0,0(Ω)Y_{0,0}(\Omega) spherical harmonic is spherically-symmetric, while Y1,0(Ω)Y_{1,0}(\Omega) is symmetric with respect to rotations about the zz-axis, and so on. The coupling of angular channels, which is embedded into the Gaunt coefficients, can be derived in a simple analytic form precisely due to the separability of these symmetry classes. This reduces the number of continuous coordinates represented numerically, as suggested by the Wigner-Eckart theorem.

Questions

Rayleigh quotient and quadrature

Two-body calculation. We use quadrature on a non-uniform grid, not importance weights. The grid is generated by sampling NN points from q(r)∝∣ψWKB(r,EWKB)∣2q(r) \propto |\psi_{\text{WKB}}(r,\mathcal{E}_{\text{WKB}})|^2. After sorting, every integral is computed via the trapezoid rule on these points, with weights wi=(ri+1−ri−1)/2w_i = (r_{i+1} - r_{i-1})/2. The optimized quantity is:

E=∑iwi[ℏ22μ∣ψ′(ri)∣2+ψ(ri)⊤Veff(ri),ψ(ri)]∑iwi,∣ψ(ri)∣2,E = \frac{\sum_i w_i \big[ \tfrac{\hbar^2}{2\mu} |\boldsymbol{\psi}'(r_i)|^2 + \boldsymbol{\psi}(r_i)^\top V_{\text{eff}}(r_i)\\,\boldsymbol{\psi}(r_i) \big]}{\sum_i w_i\\, |\boldsymbol{\psi}(r_i)|^2},

where ψ(r)\boldsymbol{\psi}(r) denotes the vector of reduced radial channel functions ψl,m(r)=ψl,mNN(r),ψWKB(r,EWKB)\psi_{l,m}(r)=\psi^{\text{NN}}_{l,m}(r)\\,\psi_{\text{WKB}}(r,\mathcal{E}_{\text{WKB}}), defined in Ψ(r,Ω)=∑l,mψl,m(r),Ylm(Ω)/r\Psi(r,\Omega)=\sum_{l,m}\psi_{l,m}(r)\\,Y_{lm}(\Omega)/r.

For most calculations, the grid is generated once before training from the fixed distribution q(r)q(r). The sampled points define the integration grid, while energy and normalization are evaluated by quadrature. Since the trial wave function is not explicitly normalized, its normalization integral remains in the denominator of the Rayleigh quotient. The quadrature weights account for the nonuniform point density: wi≈1/(Nq(ri))w_i\approx1/\big(Nq(r_i)\big) on average, so they play the role of inverse-density weighting. However, they are computed directly from the grid spacing, without evaluating or normalizing q(r)q(r). Energy, normalization, and excited-state orthogonalization all use these same weights. We added a clarification in Section 5.4.

For the VMC comparison requested by Reviewer 4BRZ, we sample points from a distribution proportional to the evolving wave-function density ∣ψ(r)∣2|\boldsymbol{\psi}(r)|^2, without computing its normalization. In this case, the sample average of the local energy directly estimates the Rayleigh quotient above, so no separate normalization denominator is required in the VMC estimator.

Three-body calculation. We use quadrature on a non-uniform two-dimensional grid spanned by the Jacobi radii (r,R)(r,R). Again, the sampled points only define the integration grid. We construct a Delaunay triangulation and integrate the piecewise-linear interpolant of each integrand: ∫Ωhf(r,R),dr,dR≈∑iwif(ri,Ri)\int_{\Omega_h} f(r,R)\\,dr\\,dR\approx\sum_i w_i f(r_i,R_i), where wi=13∑T∋iATw_i=\frac{1}{3}\sum_{T\ni i}A_T, ATA_T is the area of triangle TT and Ωh\Omega_h is the convex hull of the integration points. The vector ψ(r,R)\boldsymbol{\psi}(r,R) collects the reduced radial channel functions ψl,ml;L,ML2D(r,R)\psi^{\mathrm{2D}}_{l,m_l;L,M_L}(r,R). The radial Jacobian factors are absorbed through the 1/(rR)1/(rR) factor in the angular expansion, so their normalization uses the flat measure dr,dRdr\\,dR. We optimize:

\frac{ \sum_i w_i\left[ \frac{\hbar^2}{2\mu_r} \left|\partial_r\boldsymbol{\psi}(r_i,R_i)\right|^2 + \frac{\hbar^2}{2\mu_R} \left|\partial_R\boldsymbol{\psi}(r_i,R_i)\right|^2 + \boldsymbol{\psi}(r_i,R_i)^\top V_{\mathrm{eff}}(r_i,R_i)\\, \boldsymbol{\psi}(r_i,R_i) \right] }{ \sum_i w_i \left|\boldsymbol{\psi}(r_i,R_i)\right|^2 }.$$ Energy and normalization therefore use exactly the same area weights. These account geometrically for the nonuniform point density through the areas of the adjacent triangles, without an explicit $1/q$ factor or normalization of the sampling density. **Effect of the WKB spread on the wave function tail** The proxy energy sets the decay rate of the WKB envelope, but the physical decay rate depends on the energy of each bound state. We broaden the envelope so that it leaves room for an extended tail. The shared network then learns a residual for each channel. Because the residual multiplies the envelope, it can suppress the amplitude at large radii and give each state its own radial decay. We checked this directly by comparing the learned ground-state wave function with MDVR beyond the outer turning point. At $\nu=50\\,\text{MHz}$ and $\nu=500\\,\text{MHz}$, the learned wave function for $\beta=0.01$ and $\beta=0.1$ falls well below the broadened envelope and follows the MDVR tail over six to ten orders of magnitude before the curves depart. With the literal WKB envelope $(\beta=1)$, the learned tail instead decays too rapidly. For the weakly bound first excited state, our validation covers both the energy and the radial profile. At the chosen $\beta=0.01$, the seed-averaged energy error is at most 0.004% across the four tested Rabi frequencies, including $\nu=400\\,\text{MHz}$ (near threshold). Direct comparison of the channel functions with MDVR also shows agreement in the nodes and outer profiles, with larger deviations in the weaker channels. Section 3 of the General Reply provides more detail on these ablations.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Comment

We thank Reviewers for their constructive feedback. We have taken all of their concerns into account. Our answers to specific questions are provided in the replies to each Reviewer. Here, in the General Reply, we address major concerns raised by several Reviewers at once.

1. Complete three-body (N=3)(N=3) demonstration

We thank the Reviewers RBjc and 4BRZ for highlighting this important point. In discussing the impact of our study, we emphasized the applicability of our method to higher-body systems (N>2)(N>2). To provide a direct numerical evidence for this claim, we have now performed a complete three-body calculation using the same field-linked potential as in the two-body benchmarks, following the formulation previously discussed in the Appendix A.4. The new calculation:

  • includes all three pair interactions and the resulting coupling of dual-harmonic angular channels
  • learns radial channel functions that depend simultaneously on the two Jacobi radii
  • uses a clearly specified angular basis truncation with convergence evidence

In the revised Manuscript, we provide a direct benchmark of the three-body NN solver against a reference 2D DVR calculation. Both methods use the same Hamiltonian and angular basis truncation. Their differences therefore concern the radial numerical treatment within the same angular model. The results are summarized in the table:

Rabi frequency (ν\nu, MHz)NN energy (kHz)2D DVR energy (kHz)Relative discrepancy (%)
68.3128-132.1499-128.95182.4801
97.4525-259.7694-257.74610.7850
138.8706-484.4604-482.11240.4870
226.5149-1074.1037-1071.22460.2688
452.6004-3035.4255-3032.57790.0939
600.0000-4540.3407-4531.44640.1963
800.0000-6772.4221-6754.42320.2665
1000.0000-9167.3441-9145.04960.2438
1200.0000-11693.4725-11668.80700.2114
1431.1639-14750.0374-14721.95330.1908

We used dual-angular cutoff of (lmax⁡,Lmax⁡)=(10,8)(l_{\max},L_{\max})=(10,8). Our ablations show that increasing the angular basis from 145 to 190 channels, corresponding to cutoffs (lmax⁡,Lmax⁡)=(8,8)(l_{\max},L_{\max})=(8,8) and (10,8)(10,8) respectively, changed the three-body ground-state energy by at most 0.25% at the tested frequencies, indicating weak sensitivity to this final increase in angular cutoff.

Benchmark interpretation In the two-body benchmark, the baseline MDVR data is an established result for dimer energy, allowing its discrepancy with the neural calculation to serve as an accuracy measure. In contrast, the three-body 2D DVR reference represents an independent numerical calculation of the trimer energy. Its accuracy has not yet been fully established, although its results were cross-verified via harmonic expansion and correlated Gaussian codes (to be published elsewhere). Thus, its discrepancy with the neural calculation currently measures numerical agreement rather than established error. The neural estimates are systematically lower with differences below 1% over ν∈[97.4525,1431.1639] MHz\nu\in[97.4525,1431.1639]\thinspace\text{MHz} and increasing at lower ν\nu's, such as ≈2.5%\approx 2.5\% at ν≈68 MHz\nu\approx68\thinspace\text{MHz}, where 2D DVR becomes less stable. These differences cannot presently be assigned exclusively to either method. In particular, lower-than-the-reference variational energies might indicate that the neural representation captures structure that is insufficiently resolved by the 2D DVR calculation. For the purpose of this study, however, the sub-1% agreement for the broad range of Rabi frequencies demonstrates applicability of our approach to three-body molecular systems.

2. Contribution relative to existing neural few-body solvers

Reviewers RBjc and 4BRZ raise an important question regarding the relationship between our work and existing neural few-body solvers. The cited studies and our work share several established techniques: "partial-wave decomposition followed by neural parametrization and variational optimization of coupled channels". These are standard ingredients in constructing neural few-body solvers, and neither study’s contribution can be characterized by listing them alone. Assessing the added value requires considering the interaction potentials, the resulting numerical demands, and the choices through which each formulation addresses these demands.

The interaction potential is a core ingredient. It enters the variational objective itself, as well as determines the spatial structure that the solver must resolve. Changing the potential changes both the optimization problem and the requirements on the wave-function representation and numerical integration. Our principal contribution is to establish a neural variational approach for field-linked molecular bound states, addressing the numerical demands of a long-range multi-scale interaction potential. Li et al. (2025) establish neural solutions for nuclear bound states. That demonstration does not establish that the same general ingredients can successfully resolve the molecular problem considered here. Their applicability to longer-range, more intricate interaction profiles requires its own numerical demonstration, which our study provides.

In particular, Li et al. treat short-range nuclear interactions with exponentially decaying tails, whereas our field-linked potential combines short-range repulsion (shielding) with an anisotropic long-range dipolar attraction that decays algebraically as r−3r^{-3}. This long-range coupling, together with field-tunable confinement, motivates a representation and integration strategy that accommodate substantially different radial scales. Across Rabi frequencies, the resulting bound states range from spatially extended, weakly bound states to tightly localized states in deeper wells. Excited states additionally require radial nodes and can develop extended tails near their binding threshold. The solver must resolve these spatial features while retaining the angular-channel coupling generated by the anisotropic interaction.

These demands motivate a methodological distinction from Li et al. (2025). Their network directly parametrizes radial channel functions, and the variational energy is evaluated using a uniform-grid Hamiltonian. Our formulation incorporates information from the interaction potential into both the trial wave function and the spatial quadrature. The following numerical choices implement this construction:

  • A tempered WKB-inspired envelope supplies a potential-informed radial prior, reducing the burden of learning strongly varying radial amplitudes from scratch. Tempering broadens the envelope so that a proxy energy does not impose an overly restrictive decay profile.
  • A shared neural residual represents the remaining radial channel functions, including state-dependent localization, relative channel amplitudes, and excited-state nodes. This retains the flexibility needed across interaction strengths while providing a common parametrization of the coupled functions instead of independent optimization variables at every radial grid point.
  • A fixed nonuniform quadrature grid resolves both localized structure and extended tails without the cost of uniformly fine spatial resolution. The WKB distribution determines point placement, while quadrature weights account for nonuniform spacing. Fixing the grid also keeps the discrete variational objective consistent throughout optimization.

The benchmarks and ablations test the effectiveness of this construction. The two-body calculations cover ground and excited states across interaction strengths. The envelope and sampling experiments examine the numerical choices needed to optimize these states. For instance, in a two-body system, across four Rabi frequencies, the fixed WKB grid yielded ground-state energy errors of 0.001–0.010% relative to MDVR, compared with 0.5–6.2% on a uniform grid with the same number of points. A fixed log-uniform grid gave errors of 0.001–0.003% in the same test. Both nonuniform grids therefore improved substantially on uniform spacing, while the WKB construction provides a prescription for choosing point locations from the interaction potential. Lastly, the new complete three-body calculation tests jointly dependent radial functions and coupled angular channels in an actual three-molecule Hamiltonian.

The main added value is thus the demonstrated ability to solve field-linked molecular few-body problems. The WKB-inspired envelope and nonuniform quadrature are methodological developments supporting this capability.

3. Envelope sensitivity and asymptotic behavior

Reviewers RBjc and FM5g ask whether the proxy energy or empirical tempering of the WKB-inspired envelope bias the final energies or wave-function tails. Methodologically, the envelope provides a radial prior rather than fixing the exact state-dependent decay, hence the neural residual is capable of recovering the remaining structure. We have extended the original ablation (for both ground and excited states) to test this behavior directly.

In particular, we have replaced the original ablation with a thorough evaluation over a range of WKB-spread magnitudes, Rabi frequencies and random seeds, for both the ground and excited states. We also added a comparison of the learned tails with the envelope and the MDVR reference. In the main text, the training points are sampled from the envelope, thus β\beta changes both the form of the wave function and where it is trained. To isolate the effect of the envelope, for ablations, we train the network on a fixed log-spaced grid that is identical for all β\beta. We performed the optimization with β∈0,0.001,0.01,0.1,1\beta\in\\{0, 0.001, 0.01, 0.1, 1\\} (in the main text, β=0.01\beta=0.01), for the ground state at ν∈50,100,500,1000\nu\in\\{50,100,500,1000\\} MHz and for the excited state at ν∈400,500,700,1000\nu\in\\{400,500,700,1000\\} MHz, where it is bound. Each configuration was evaluated at three random seeds, with all other settings fixed. To calculate the excited state, we used the ground state of the same ν, β\nu,\,\beta and random seed for the orthogonalization projection. Note that, β=0\beta=0 removes the envelope decay and β=1\beta=1 represents the literal WKB envelope. The table below provides the seed-averaged relative error with respect to MDVR:

β\betaGround state, error vs MDVR (%)Excited state, error vs MDVR (%)
00.0001 – 0.00020.0003 – 0.001
0.0010.0001 – 0.00030.0003 – 0.001
0.010.0002 – 0.00050.0004 – 0.004
0.10.007 – 0.0110.003 – 1.3
10.03 – 2.220 – 76 ‡

‡ For ν=400\nu=400 and 500500 MHz (and β=1\beta=1), the solver for an excited state gets stuck above the dissociation threshold (E>0)(E>0), hence the presented value is averaged over the two remaining Rabi frequencies, where the solver converges to E<0E<0.

For the ground state, the accuracy is insensitive to β\beta over two orders of magnitude, and all runs with β≤0.1\beta\le0.1 agree with MDVR to within about 0.01%. In this range the learned wave function decays far below the envelope and follows the MDVR tail over many orders of magnitude, so the network recovers the correct asymptotic decay rather than inheriting it from the prior. The excited state lies much closer to threshold and is more sensitive, but β≤0.01\beta\le0.01 still agrees with MDVR to within about 0.005% at all frequencies. The spread mainly affects convergence. A moderate spread (β=0.01\beta=0.01 and 0.10.1) brings the energy to the MDVR value within a few hundred steps, while without a decaying envelope the energy first stays near the threshold and converges several times more slowly. The literal envelope (β=1\beta=1) is too restrictive. Its learned tail falls off faster than the reference, the ground-state error grows towards low frequency, and the excited state fails at every frequency. The value β=0.01\beta=0.01 used in the paper therefore lies within the range where both states are accurate and converge quickly. The full results per frequency and seed, with training curves and the tail comparison, are in Appendix E.3.1.

All new results mentioned in this General Reply, as well as in the per-Reviewer comments, will be included in the revised Manuscript, which will be made available before the end of the discussion period (Nov 18, 2026). For clarity, all major changes in the revised Manuscript are marked in red.

We have uploaded the revised Manuscript. Major changes are marked in red. Section 7 now presents a complete field-linked three-body calculation and compares its ground-state energies with an independent 2D DVR calculation. Appendix E.3 addresses the Reviewers’ questions about envelope sensitivity and wave-function tails, angular-cutoff convergence, sampling strategy, and independent evaluation of the trained networks. The updated two-body benchmarks are in Section 6.1, with five-seed results in Appendix E.4. We thank the reviewers for the suggestions that motivated these revisions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.