A Boson-informed neural network for Ground States of Many-Body Schr" o dinger equations

arXiv:2610.18875 · physics.comp-ph, cond-mat.quant-gas, math-ph, math.MP · Submitted 2026-09-05 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: Today's paper: "A Boson-informed neural network for Ground States of Many-Body Schr" o dinger equations".

Mira: The gist: A symmetry-preserving Boson-informed neural network (BINN) for continuous space computation of interacting-boson ground states has been proposed,

Kai: First, who's behind it and why it matters.

Paper summary: Kai: Hey Mira, so we're looking at this paper today called "A Boson-informed neural network for Ground States of Many-Body Schrodinger equations" <ref:2610.18875#pg1>. It's about using a symmetry-preserving Boson-informed neural network, or BINN, to do variational Monte Carlo computation for interacting boson ground states in continuous space <ref:2610.18875#pg1>. Basically, they're trying to solve the tough problem of finding these ground states where the particles are actually interacting with each other <ref:2610.18875#pg2>.

Mira: Yeah, what's really interesting here is how they split up the physics. The core idea is to explicitly separate the two main physical scales of the system <ref:2610.18875#pg1>. They put a cell Jastrow factor on top to handle that microscopic interaction scale, which deals with things like near-contact stiffness <ref:2610.18875#pg4>.

Lev: So, what does that mean for the actual computation? I'm thinking about how much overhead that adds <ref:2610.18875#pg3>. If we have to factor in these components, how does that affect running this on actual hardware <ref:2610.18875#pg3> and where does the complexity land in terms of quantum error correction needs <ref:2610.18875#pg3>?

Kai: Well, they claim that by separating the scales, they leave the neural network to just capture a smoother many-body residual <ref:2610.18875#pg1>. They also show that across both two-dimensional and three-dimensional tests, this BINN method gets close to the Hartree and Gross-Pitaevskii limits <ref:2610.18875#pg4>.

Mira: And they give some concrete numbers for how good it is, saying they see an empirical O(N-one) finite-size trend in both energy and density <ref:2610.18875#pg1>. They also mention that the energies agree with diffusion Monte Carlo to within zero point zero six percent for the particle number tested up to three hundred fifty in the three-dimensional shrinking-range case <ref:2610.18875#pg4>.

Lev: A zero point zero six percent error rate is pretty tight when you're talking about these complex systems, especially if we're trying to scale it up for real hardware <ref:2610.18875#pg3>. But I wonder about the way they handle the confinement part, because that’s where things usually get messy <ref:2610.18875#pg4>.

Kai: Right, that’s exactly what they do by having a potential-aware envelope <ref:2610.18875#pg4>. They use this to encode the external confinement structure analytically, which lets the neural network focus only on learning that smoother many-body part <ref:2610.18875#pg4>.

Paper summary: Mira: The architecture they propose is based on a Symmetry-Jastrow factorization for trapped interacting bosons <ref:2610.18875#pg4>. This factorization splits the system into two branches, one for the one-body part and one for the pair part <ref:2610.18875#pg4>.

Lev: So, if you're looking at a real quantum computer setup, that means they're building a specific structure to preserve permutation symmetry <ref:2610.18875#pg4>. That’s important for keeping the state physically correct <ref:2610.18875#pg4>.

Kai: Exactly, and they use a tensor neural network, or TNN, to build that one-body backbone <ref:2610.18875#pg4>. This TNN has to be positive and permutation-symmetric so the whole BINN stays symmetric <ref:2610.18875#pg5>.

Mira: They also use this cell Jastrow factor which is defined by a function u that resolves the contact stiffness in one dimension <ref:2610.18875#pg4>. And they use a smooth neural network as a residual to learn the regular part of those correlations <ref:2610.18875#pg4>.

Lev: That sounds like they are trying to make the local energy evaluation closed-form, which is smart for speeding things up <ref:2610.18875#pg3>. But what about the training part? How do you actually get those parameters theta tuned correctly <ref:2610.18875#pg6>?

Kai: They use a Stochastic Reconfiguration optimizer, or SR, which they say is a standard natural-gradient optimizer for neural quantum states <ref:2610.18875#pg6>. It updates the parameters by solving a P times P linear system defined by the Fisher-information matrix S <ref:2610.18875#pg6>.

Mira: The training process itself involves sampling configurations R from the probability density rhotheta using the Metropolis-adjusted Langevin algorithm, or MALA <ref:2610.18875#pg5>. After sampling, they evaluate the local energy EL(R) using some specific formulas <ref:2610.18875#pg5>.

Lev: And then they take one SR step to get the next set of parameters theta(t+one) <ref:2610.18875#pg5>. So it’s an iterative loop between sampling and updating, which is standard stuff in this field <ref:2610.18875#pg3>.

Kai: Looking at the validation, they tested this across various trapping scenarios, including ring-shaped, lattice, double-well, and box traps <ref:2610.18875#pg20>. They show robust performance in all of them <ref:2610.18875#pg4>.

Paper summary: Mira: And the ablation studies they did were pretty telling <ref:2610.18875#pg3>. They showed that removing either the cell Jastrow factor or the potential-aware envelope makes both EVMC and CV worse <ref:2610.18875#pg18>. That really proves why those components are necessary for the performance <ref:2610.18875#pg4>.

Lev: So, they can't just throw away either the contact stiffness absorber or the confinement encoder and expect it to work <ref:2610.18875#pg4>. That tells us that the separation of scales is fundamental for this method <ref:2610.18875#pg4>.

Kai: It means they can achieve state-of-the-art accuracy, approaching the Hartree and GrossPitaevskii limits <ref:2610.18875#pg5>. They even maintain DMC-level accuracy when the system size grows and the interaction range shrinks <ref:2610.18875#pg4>.

Mira: That's what's compelling, because maintaining that level of accuracy across different particle numbers and shrinking interaction ranges is a big deal for practical applications <ref:2610.18875#pg5>.

Lev: The limitation they point out is that this method isn't just about the ground state calculation itself <ref:2610.18875#pg3>. They also say that the BINN's architecture needs to be robustly transferable across different confinement potentials <ref:2610.18875#pg3>. That means proving it works well in scenarios not explicitly tested <ref:2610.18875#pg20>.

Kai: So, the implication for us is that this BINN framework is a novel and robust way to compute ground states for interacting bosons in continuous space <ref:2610.18875#pg6>. It’s designed to separate microscopic interaction stiffness via a cell Jastrow factor from macroscopic confinement encoded in a potential-aware envelope <ref:2610.18875#pg4>.

Mira: Exactly, and the paper concludes that this architectural design allows the neural network to capture the smoother many-body residual <ref:2610.18875#pg4>. It’s a novel framework for variational Monte Carlo computation of ground states for interacting bosons in continuous space <ref:2610.18875#pg6>.

Lev: So, the paper shows that by separating those two dominant physical scales, you can get very high accuracy in the ground state energy calculation <ref:2610.18875#pg4>. It’s a solid foundation for how we might approach these problems on actual quantum hardware <ref:2610.18875#pg3>.

Kai: Yeah, it’s a novel and robust framework that actually works across different trapping potentials <ref:2610.18875#pg20>. It shows that BINN is a solid way to go for these continuous space problems <ref:2610.18875#pg6>.

Conclusion: Kai: So, we’re wrapping up this look at "A Boson-informed neural network for Ground States of Many-Body Schrodinger equations" <ref:2610.18875#pg6>. Basically, they built a way to use a neural network to find the ground states for interacting bosons in continuous space <ref:2610.18875#pg6>.

Mira: It’s pretty neat how they structured it by separating the physical scales, using those Jastrow factors and potential envelopes to handle microscopic stiffness and macroscopic confinement separately <ref:2610.18875#pg4>.

Lev: From my side, I'm looking at how this separation helps us with error correction; if we can isolate the physics like that, maybe the overhead for simulating it on a real computer gets lower <ref:2610.18875#pg3>.

Kai: Yeah, and they show that this approach gets very close to what we expect from simpler theories like Hartree or Gross-Pitaevskii limits <ref:2610.18875#pg4>.

Mira: That's the big picture idea, getting near those known limits even with the neural network part handling the residual stuff <ref:2610.18875#pg4>.

Lev: But they do have some constraints; they mentioned that while it works well for certain traps, you need to test it across different scenarios like ring-shaped or lattice traps <ref:2610.18875#pg20>.

Kai: Exactly, and the ablation studies showed that you can't just remove either the Jastrow factor or the envelope; removing either one hurts the accuracy <ref:2610.18875#pg18>.

Mira: So, they’re arguing that having both components is necessary for it to work well across all these different physical setups <ref:2610.18875#pg4>.

Lev: And that leads us to thinking about the future; if this separation of scales holds up, it means we might be able to use these AI models as better tools for finding ground states in much more complex continuous systems <ref:2610.18875#pg6>.

ZHIZHONG KONG, JERRY ZHIJIAN YANG, CHENG YUAN, XIAOFEI ZHAO

Wuhan University

physics.comp-ph, cond-mat.quant-gas, math-ph, math.MP

Submitted: 2026-09-05

Updated: 2026-09-05

Code: https://github.com/kongzhizhong/Boson-informed-Neural-Network

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: The gist: A symmetry-preserving Boson-informed neural network (BINN) for continuous space computation of interacting-boson ground states has been proposed, which explicitly separates physical scales

Key concepts

Cell Jastrow Factor: This analytical component is used to resolve the leading near-contact stiffness in the microscopic interaction scale of the system. It is constructed to absorb singularities in specific terms, allowing a smooth neural network to focus on learning the regular part of the correlation effects between bosons.
Potential-Aware Envelope: This term analytically encodes the external confinement structure of the system. By incorporating this envelope into the trial state, it allows the main neural network backbone to concentrate its learning efforts solely on capturing the smoother many-body residual behavior, simplifying its task.
Symmetry-Jastrow Factorization: This universal architecture is based on a proposed factorization that preserves permutation symmetry. It splits the system's wave function into branches—one for single particles and one for pairs—which together integrate the Jastrow factor and the potential envelope into a structure suitable for local energy evaluation.
Symmetry-Preserving Neural Network (TNN): The neural backbone is constructed using a positive, permutation-symmetric tensor neural network. This specific architecture ensures that the overall Boson-informed neural network maintains the necessary permutation symmetry required for describing interacting bosons accurately.

Summary

The gist: A symmetry-preserving Boson-informed neural network (BINN) for continuous space computation of interacting-boson ground states has been proposed, which explicitly separates physical scales into a cell Jastrow factor and a potential-aware envelope to capture microscopic interactions and macroscopic confinement, respectively.

How it works

The BINN framework is designed with a universal architecture based on a proposed Symmetry-Jastrow factorization for trapped interacting bosons, preserving permutation symmetry<ref:2610.18875#pg4>. This factorization separates the system into two branches: a one-body branch and a pair branch<ref:2610.18875#pg4>. The two branches integrate a cell Jastrow factor, which resolves the contact stiffness, and a potential-aware envelope, which captures the macroscopic confinement background, into a pooled structure that allows for closed-form local-energy evaluation<ref:2610.18875#pg4>.

The full trial state is constructed as Ψθ(R) = Φsymθ1(R) ΦJasθ2(R), where the Jastrow component is defined as exp X i. The symmetric wave function Φsymθ1 is a log-space construction composed of a potential-aware one-body envelope sum and a pooled neural backbone<ref:2610.18875#pg4>. The neural backbone fTNNθ1 is constructed from a positive, permutation-symmetric tensor neural network (TNN), which ensures the full BINN (3.1) is permutationsymmetric<ref:2610.18875#pg5>.

Key Components and Architecture

The BINN architecture separates the system's dominant physical scales by incorporating specific analytical components<ref:2610.18875#pg4>. The cell Jastrow factor aims to capture correlation effects between bosons through the key 1D function u: [0,∞) → R, which is written as u = uθ2 for brevity<ref:2610.18875#pg10>. This factor is constructed analytically to absorb the singularity in vσ, and a smooth neural network is then employed as a residual to learn the regular part of the correlations<ref:2610.18875#pg4>.

The potential-aware one-body envelope χenvθ1 is introduced via (3.2) to analytically encode the confinement structure, allowing the neural backbone to focus solely on refining the remaining many-body behavior<ref:2610.18875#pg10>. This envelope is defined as χenvθ1(r) = −X L γ=1 αγvγ(r), where all coefficients are treated as trainable parameters<ref:2610.18875#pg10>.

The local energy ELΨθ is evaluated by decomposing the kinetic term into exact one-body and two-body parts, and the neural term which is the Laplacian of fTNNθ1: R dN → R<ref:2610.18875#pg13>. This allows for efficient evaluation by replacing the calculation of a dN-dimensional function's Laplacian with R different d-dimensional functions evaluated at N points<ref:2610.18875#pg14>.

Optimization and Training

The parameters θ are updated using the Stochastic Reconfiguration (SR) optimizer, which is a standard natural-gradient optimizer for neural quantum states<ref:2610.18875#pg6>. The energy gradient is estimated by ∇θE[Ψθ] ≈ 2 B O⊤ c e<ref:2610.18875#pg6>. The SR update solves a P × P linear system (2.8) to update θ, where S is the Fisher-information matrix<ref:2610.18875#pg6>.

The training process involves drawing configurations R from ρ[Ψθ] using the Metropolis-adjusted Langevin algorithm (MALA)<ref:2610.18875#pg5>. After configuration sampling, the local energy EL(R(s)) is evaluated using (2.5) and (3.13)<ref:2610.18875#pg5>. The state is then updated by one SR step (2.8) for θ(t+1)<ref:2610.18875#pg5>.

Numerical Validation and Results

The BINN method was validated across various tests, showing accuracy, scalability, and asymptotic consistency with theoretical results<ref:2610.18875#pg4>. In the fixed-range regime (2D harmonic trap), the relative energy gap between BINN and DMC is below 0.06% for all N<ref:2610.18875#pg5>.

In the shrinking-range regime (3D), BINN maintains Diffusion Monte Carlo level accuracy for particle numbers up to 350 in three dimensions, with a relative energy discrepancy as low as 0.008%<ref:2610.18875#pg5>. The method scales to N = 350 in the threedimensional shrinking-range regime (σ ≈ 0.03) while maintaining DMC-level accuracy<ref:2610.18875#pg4>.

A series of ablation studies demonstrated the necessity of the cell Jastrow factor and the potential-aware envelope, showing that removing any key component leads to an increase in both EVMC and ΓCV, thereby degrading overall performance<ref:2610.18875#pg18>. Applications across various trapping potentials—including ring-shaped, lattice, double-well, and box traps—underscore the necessity of our proposed architectural components and the robust generalizability of BINN to diverse physical scenarios<ref:2610.18875#pg20>. This demonstrates that BINN achieves state-of-the-art accuracy, approaching the Hartree and GrossPitaevskii limits<ref:2610.18875#pg5>. The method maintains DMC-level accuracy even as the system size grows and the interaction range shrinks<ref:2610.18875#pg5>.

The BINN framework successfully separates microscopic interaction stiffness via a cell Jastrow factor from macroscopic confinement encoded in a potential-aware envelope to capture the smoother many-body residual<ref:2610.18875#pg4>. This architectural design allows the neural network to capture the smoother many-body residual<ref:2610.18875#pg4>. The paper concludes that BINN is a novel and robust framework designed for variational Monte Carlo computation of ground states for interacting bosons in continuous space<ref:2610.18875#pg6>. This framework effectively separates the system’s dominant physical scales through a Symmetry-Jastrow factorization, separating microscopic interaction stiffness via a cell Jastrow factor from macroscopic confinement encoded in a potential-aware envelope<ref:2610.18875#pg4>. This architectural design allows the neural network to capture the smoother many-body residual<ref:2610.18875#pg4>. The paper concludes that BINN is a novel and robust framework designed for variational Monte Carlo computation of ground states for interacting bosons in continuous space<ref:2610.18875#pg6>.

--- Page 26 ---

The paper is 2610.18875. EVERY sentence must end with a reference token<ref:2610.18875; the token names the page and quotes the words it rests on, copied word for word from that page).

How it works

The full trial state is constructed as Ψθ(R) = Φsymθ1(R) ΦJasθ2(R), where the Jastrow component is defined as exp X i.

Improvements for AI systems

Boson-informed neural network (BINN) architecture for ground state computation: This system can compute interacting-boson ground states in continuous space by explicitly separating physical scales: At the microscopic interaction scale, a cell Jastrow factor absorbs the leading near-contact stiffness, while at the macroscopic scale, a potential-aware envelope encodes the external confinement.
Near-DMC accuracy in shrinking-range regimes: The BINN can maintain Diffusion Monte Carlo level accuracy for particle numbers up to 350 in three dimensions, even as interaction range shrinks, achieving a relative discrepancy as low as 0.008%.
Robustness across diverse trapping potentials: The system is generalizable because it handles various traps, including ring-shaped, deep-lattice, double-well, and box traps, demonstrating that the architecture inherently encodes the trapping environment.
Scalable training and evaluation: The implementation utilizes a Stochastic Reconfiguration (SR) optimizer to update parameters via a linear system: (S + λI) δθ = −η ∇θE[Ψθ], enabling training on systems up to N=350 with practical memory usage, validated by the Pooled calculation method.
Asymptotic convergence analysis: The system can provide quantitative benchmarks by showing trends toward mean-field limits, as the computed GS shows a trend O(N −1) toward mean-field limits in both fixed-range and shrinking-range regimes.

Sources

Neural network quantum states in the grand canonical ensemble