🔬 mesoscale physics

Scaling universal Fermi network toward ground states: A diffusion-Monte-Carlo assessment

This paper demonstrates that the universally expressive Fermi Sets neural network architecture can be systematically scaled to converge to the ground states of interacting fermionic systems, as evidenced by the diminishing energy improvements from fixed-phase diffusion Monte Carlo calculations on jellium and quantum dot models.

Original authors: Yu-Sheng Li, Saskia Poldmaa, Tzen Ong, Ahmed Abouelkomsan, Tay-Rong Chang, Hsin Lin, Liang Fu

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yu-Sheng Li, Saskia Poldmaa, Tzen Ong, Ahmed Abouelkomsan, Tay-Rong Chang, Hsin Lin, Liang Fu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict the exact dance of a crowd of invisible, jittery dancers. In the world of quantum physics, these dancers are electrons, and the choreography they follow is dictated by a set of rules called the Schrödinger equation. This equation is the "script" that tells us how atoms, molecules, and the materials around us behave. But here's the catch: electrons are tricky. They are "antisocial," meaning if you swap two of them, the whole dance routine flips upside down (a rule called antisymmetry). Furthermore, the stage they dance on is infinite, and the number of possible moves they can make is so vast it's practically impossible to count.

For decades, scientists have tried to solve this dance script using two main strategies. The first is like trying to memorize every single step in a library of books; it's accurate but gets so huge and slow that it breaks down for anything more complex than a few dancers. The second strategy is like using a rough sketch of the dance to guess the moves; it's faster, but the sketch might be wrong in ways you can't easily see. Recently, a new tool has entered the stage: artificial intelligence (neural networks). These are computer programs that can learn to write the dance script themselves, getting better and better the more "brainpower" (parameters) you give them. The big question, however, has been: how do we know if the AI has actually found the perfect dance, or just a really good guess?

This paper introduces a clever two-step method to answer that question, acting like a rigorous "final exam" for these AI dancers. The researchers used a specific type of AI architecture called "Fermi Sets," which is designed from the ground up to respect the antisocial rules of electrons. They first let the AI practice and minimize its energy (finding the most efficient dance). Then, they ran a second, more powerful simulation called "Diffusion Monte Carlo" (DMC) on top of the AI's result. Think of this second step as a strict coach who takes the AI's dance and tries to tweak it just a tiny bit to see if it can get any better. If the coach can't improve the dance at all, it means the AI has likely found the true, perfect ground state.

The team tested this on two very different scenarios. First, they looked at a "quantum dot," which is like a tiny, artificial atom where electrons are trapped in a magnetic field. In this setting, the electrons perform a complex, swirling dance that requires a "phase" (a hidden direction in the math) that is hard to get right. They found that as they made the AI network bigger and smarter, the difference between the AI's guess and the coach's final check shrank to almost nothing—so small it was practically zero. In fact, their AI found a specific, exotic dance pattern (a composite fermion state) that traditional methods couldn't even describe.

Next, they applied the same method to a "jellium" model, which is like a uniform sea of electrons floating in a neutral background, representing a metal. Here, the electrons are strongly interacting and crowded. Again, as they increased the size of the neural network, the "gap" between the AI's energy and the coach's energy collapsed. The paper shows that for both the trapped quantum dot and the electron sea, the AI's performance systematically improved until it was indistinguishable from the exact solution. The researchers conclude that by scaling up these "Fermi Sets" networks, they can systematically approach the exact ground state of complex electron systems, offering a powerful new way to solve some of the hardest problems in quantum physics.

Technical Summary: Scaling Universal Fermi Network toward Ground States

Problem Statement
Solving the electronic Schrödinger equation for interacting many-electron systems remains a central challenge in condensed matter physics and quantum chemistry. The difficulty arises from the infinite-dimensional Hilbert space of continuum systems and the requirement that fermionic wavefunctions be antisymmetric under particle exchange. Traditional methods face significant limitations: Full Configuration Interaction (FCI) scales factorially with system size and is restricted by finite basis sets; truncated expansions (e.g., Coupled Cluster) suffer from both basis set incompleteness and many-body truncation errors; and Diffusion Monte Carlo (DMC), while operating in continuous space, relies on the fixed-node approximation for real-valued trial states, introducing uncontrolled errors if the nodal structure is imperfect.

While recent neural network approaches (e.g., FermiNet, PauliNet) have shown promise by directly generating antisymmetric wavefunctions without basis sets, a critical gap remains: there is no rigorous, systematic method to assess how close an optimized neural wavefunction is to the exact ground state, particularly when it outperforms existing benchmarks. Furthermore, the feasibility of scaling these networks to achieve asymptotically exact results has not been fully explored.

Methodology
The authors introduce a two-stage computational framework combining Fermi Sets—a provably universal neural network architecture for fermionic wavefunctions—with Fixed-Phase Diffusion Monte Carlo (FPDMC).

Fermi Sets Architecture (Variational Optimization):
The wavefunction is constructed as a linear combination of a small number ( $N_d$ ) of Slater determinants with learnable complex-valued orbitals, weighted by symmetric, permutation-invariant neural networks (e.g., Deep Sets or Transformers). The ansatz is given by:
$\Psi_\theta(R) = e^{J_\theta(R)} e^{E(R)} \sum_{k=1}^{N_d} \Omega_k(R) \det[\phi^k_i(r_j)]$
where $J_\theta$ is a Jastrow factor (handling electron-electron cusps), $E(R)$ is an optional envelope (ensuring normalizability in confined systems), and $\Omega_k(R)$ are symmetric functions capturing correlations. This architecture is proven to be a universal approximator for continuous antisymmetric functions. The network is optimized via Variational Monte Carlo (VMC) to minimize the energy.
Fixed-Phase Diffusion Monte Carlo (Assessment and Refinement):
The optimized neural wavefunction $\Psi_\theta(R) = |\Psi_\theta(R)| e^{i\Phi_\theta(R)}$ serves as the trial state for an imaginary-time FPDMC projection. In this method, the phase $\Phi_\theta(R)$ is fixed, and the amplitude evolves under an effective Hamiltonian that includes a repulsive potential derived from the phase gradient.
- Metric for Accuracy: The framework utilizes the energy gap $\Delta E = E_{\text{VMC}} - E_{\text{DMC}}$ . Since $E_{\text{DMC}}$ is the variational lower bound for the specific phase manifold learned by the network, a vanishing $\Delta E$ indicates that the amplitude error of the neural network has been minimized to the limit allowed by its phase.
- Convergence Criterion: If $\Delta E \to 0$ as the network capacity increases, it implies the network has converged to the true ground state (assuming the learned phase is correct).

Key Results
The framework was applied to two distinct systems to demonstrate scalability and convergence:

Quantum Dot in a Magnetic Field:
- System: $N=4$ electrons in a 2D harmonic trap with GaAs parameters under magnetic fields $B=4$ T and $B=8$ T.
- Findings: The ground states are intrinsically complex-valued (breaking time-reversal symmetry), corresponding to composite fermion states (e.g., $\nu=3/7$ at 8 T).
- Scaling: As the number of network parameters increased (up to $8.4 \times 10^4$ ), the variational energy ( $E_{\text{VMC}}$ ) decreased monotonically, surpassing multi-Landau-level exact diagonalization (ED) benchmarks.
- Convergence: The residual gap $\Delta E$ collapsed monotonically from $\sim 561 \mu H^*$ to $\sim 25 \mu H^*$ (at 8 T) and $\sim 24 \mu H^*$ (at 4 T), becoming statistically indistinguishable from zero. This confirms the network found the exact ground state within statistical resolution.
Two-Dimensional Uniform Electron Gas (2D UEG):
- System: $N=16$ fully spin-polarized electrons at $r_s = 30$ (deep in the strongly correlated regime, near Wigner crystallization).
- Findings: The network was scaled from $\sim 25,000$ to $\sim 50,000$ parameters.
- Scaling: Both $E_{\text{VMC}}$ and the projected $E_{\text{DMC}}$ decreased monotonically. The lowest energy achieved ( $E_{\text{DMC}} \approx -0.031996$ Ha) improved upon the best Slater-Jastrow benchmarks.
- Convergence: The residual gap $\Delta E$ decreased systematically from $1.63 \times 10^{-5}$ to $0.57 \times 10^{-5}$ Ha per electron as capacity grew, indicating systematic suppression of amplitude error.

Significance and Claims
The paper claims that Fermi Sets, when combined with FPDMC assessment, provides a universal and asymptotically exact solver for the electronic Schrödinger equation. The primary significance lies in the demonstration that:

Neural network wavefunctions can be systematically scaled to approach the exact ground state, with the variational energy converging as network capacity increases.
The fixed-phase DMC gap ( $\Delta E$ ) serves as a rigorous, quantitative metric to certify that a neural wavefunction has saturated its variational freedom and converged to the ground state for a given phase manifold.
The architecture successfully captures complex nodal structures (such as composite fermion states) that are inaccessible to traditional Slater-Jastrow ansätze, which cannot describe the necessary complex phase winding.

The authors conclude that while the phase error is not directly probed by $\Delta E$ , the monotonic collapse of the gap with network size provides strong evidence that Fermi Sets can achieve arbitrary accuracy for many-electron ground states, limited only by network size and computational resources.

Technical Summary: Scaling Universal Fermi Network toward Ground States

More like this