⚛️ quantum physics

Foundation Neural Effective Hamiltonian for Strongly Correlated Quantum Materials

The paper introduces the Foundation Neural Effective Hamiltonian (FNEH), a method that projects families of Hamiltonians onto a compact subspace of foundation neural quantum states to enable accurate, low-cost exploration of phase boundaries and competing phases in strongly correlated quantum materials without requiring repeated neural-network sampling at every target coupling.

Original authors: Lixing Zhang, Hongjie Jiang, Di Luo

Published 2026-08-17

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Lixing Zhang, Hongjie Jiang, Di Luo

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 how a crowd of people will move through a city. If you only watch one specific street corner, you might get a good idea of the flow there. But what if you need to understand the entire city as the weather changes, the time of day shifts, or new roads open up? In the world of quantum physics, scientists face a similar challenge. They study "strongly correlated materials," which are like super-complex crowds of electrons that interact with each other in wild, unpredictable ways. To understand these materials, physicists use something called a "Hamiltonian," which is essentially a giant mathematical recipe that tells them how the electrons behave. Usually, they have to solve this recipe from scratch for every single tiny change in the environment, like adjusting the pressure or the temperature. It's like trying to write a new, perfect story for every single page of a book, which takes forever and costs a lot of computing power.

Recently, scientists have tried using "Foundation Neural Quantum States" (FNQS). Think of this as training a super-smart AI to learn the general rules of how these electron crowds behave across many different conditions at once. It's like teaching an AI to write a story that can adapt to different weather or settings without needing a full rewrite. However, this AI sometimes gets confused when the story takes a sharp turn, like when the material suddenly changes from a solid to a liquid (a phase transition). It might miss the exact moment the change happens or guess the wrong outcome. Furthermore, even though the AI is smart, checking its work for every single new condition still requires a massive amount of computer time, like asking the AI to re-read its own story for every new chapter.

This is where a new method called the Foundation Neural Effective Hamiltonian (FNEH) comes in, introduced by researchers Lixing Zhang, Hongjie Jiang, and Di Luo. Instead of just asking the AI to guess the answer for every new condition, FNEH takes the AI's best guesses from a few key locations and uses them to build a compact, reusable "map" of the entire landscape. Imagine you have a few high-quality photos of a mountain taken from different angles. Instead of trying to guess what the whole mountain looks like from a single photo, you combine those photos to create a 3D model. Once you have that model, you can instantly see what the mountain looks like from any angle, or even predict what happens if it snows, without taking a single new photo.

The paper demonstrates that this method works incredibly well for "moiré materials," which are special layers of semiconductors that create a wavy, patterned landscape for electrons. The researchers trained their AI on a specific range of conditions and then used FNEH to explore the space between those points. They found that while the original AI sometimes got the "phase transition" wrong—predicting the electrons were in a solid, crystalline pattern when they were actually flowing like a liquid—FNEH fixed this mistake. By combining the AI's different guesses, FNEH correctly identified the exact moment the material changed its state. In their simulations, this method not only found the correct physics but also slashed the computing time required. Once the initial "map" was built, scanning through hundreds of new conditions took almost no extra time, whereas the old methods would have had to start over for each one. The researchers suggest this approach opens a new, efficient path for studying complex quantum materials, allowing scientists to explore vast families of materials without getting bogged down by the sheer cost of calculation.

Technical Summary: Foundation Neural Effective Hamiltonian for Strongly Correlated Quantum Materials

Problem Statement
Simulating strongly correlated quantum materials often requires solving not a single Hamiltonian, but a family of Hamiltonians parameterized by tunable couplings (e.g., interaction strength, pressure, doping, or external fields). While Neural Quantum States (NQS) have emerged as accurate variational ansatzes, and Foundation Neural Quantum States (FNQS) offer a route to amortize calculations across parameter manifolds by training a single model on a range of couplings, they face two critical limitations:

Accuracy Degradation: FNQS predictions can become fragile near phase transitions, level crossings, or regions with sharp changes in correlation patterns, often misidentifying phase boundaries.
Sampling Costs: Evaluating multiple target couplings with FNQS typically requires repeated Monte Carlo sampling for each point, leading to computational costs that grow linearly with the number of target Hamiltonians.

Conversely, traditional subspace methods like Eigenvector Continuation (EC) construct compact subspaces from eigenstates to reconstruct nearby states efficiently. However, constructing these subspaces usually requires expensive, independent many-body calculations for each basis state, which does not scale well for large systems or dense parameter scans.

Methodology: Foundation Neural Effective Hamiltonian (FNEH)
The authors introduce the Foundation Neural Effective Hamiltonian (FNEH), a framework that projects a family of Hamiltonians onto a compact subspace spanned by FNQS wavefunctions sampled at selected couplings. The method operates as follows:

Foundation Model Training: An FNQS is trained over a parameter manifold $\Lambda_{train}$ using a loss function that minimizes the energy expectation value across selected training points. The architecture (based on Psiformer) embeds Hamiltonian parameters as features alongside electron coordinates to output conditional wavefunction amplitudes $\Psi(\lambda, R)$ .
Subspace Construction: Instead of optimizing a new wavefunction for every target coupling, FNEH selects a set of $N_{state}$ basis states $\{|\Psi(\lambda_i)\rangle\}$ from the pretrained FNQS at specific parameter points.
Effective Hamiltonian Formulation: For a Hamiltonian linear in parameters, $\hat{H}(\lambda) = \sum_k \lambda_k \hat{h}_k$ , the method constructs parameter-independent matrices $\hat{A}_k$ :
$\hat{A}_k = \hat{S}^{-1} \tilde{h}_k$
where $\tilde{h}_k$ and $\hat{S}$ are the operator and overlap matrices, respectively, computed between the basis states. The effective Hamiltonian for any target $\lambda$ is then reconstructed as $\hat{H}_{FNEH}(\lambda) = \sum_k \lambda_k \hat{A}_k$ .
Determinant-State Sampling (DSS): To compute the matrix elements efficiently and avoid the statistical noise amplification associated inverting the overlap matrix $\hat{S}$ , the authors employ Determinant-State Sampling. This involves sampling extended electron configurations from the probability distribution of the determinant of the basis state matrix, allowing for the stable estimation of $\hat{A}_k$ .
Variational Improvement: Once the matrices are sampled, the ground state for any target $\lambda$ is found by diagonalizing the small $N_{state} \times N_{state}$ matrix $\hat{H}_{FNEH}(\lambda)$ . This variationally combines the FNQS basis states to improve the approximation beyond any single basis state.

Key Contributions

Hybrid Framework: FNEH integrates the transferability of foundation models with the variational accuracy of subspace projection methods. It requires no additional target-specific optimization; the basis states are generated purely by inference from the pretrained FNQS.
Phase Transition Resolution: By variationally combining basis states from different parameter regions, FNEH can recover ground-state branches and phase boundaries that the underlying foundation model misidentifies or smooths over.
Computational Efficiency: The method decouples the expensive sampling step (performed once to build $\hat{A}_k$ ) from the parameter sweep. Subsequent evaluations of couplings, observables, and phase boundaries scale only with the small dimension of the effective Hamiltonian, rather than the system size or the number of target points.

Results
The authors demonstrate FNEH on a continuum model of interacting electrons in a moiré potential, a platform known for rich strongly correlated and topological phases.

Accuracy: In benchmarking against an "optimal reference" (separately optimized NQS for each point) and Exact Diagonalization (ED), FNEH significantly outperforms the raw FNQS. Specifically, in a regime where FNQS incorrectly predicts a Charge Density Wave (CDW) phase, FNEH correctly recovers the Fermi Liquid (FL) phase and the correct phase boundary location (e.g., near $\phi \approx 40.2^\circ$ ).
Energy Lowering: FNEH systematically lowers the variational energy compared to the single FNQS prediction, with improvements up to $\approx 0.1$ meV per electron in tested regions.
High-Resolution Scans: The method enables rapid, high-resolution multidimensional phase scans (e.g., over potential strength $V_0$ and phase $\phi$ ). The resulting potential energy surfaces clearly resolve sharp CDW-to-FL transitions that appear smeared in the raw FNQS predictions.
Cost Reduction: Computational timing analysis shows that while the cost of FNQS and optimal reference methods grows linearly with the number of target states, FNEH cost remains approximately constant after the initial matrix construction. FNEH becomes more efficient than FNQS when scanning more than $\approx 12$ target states.

Significance
The paper claims that FNEH opens a new avenue for studying strongly correlated quantum materials using foundation models. By transforming FNQS from a parameter-conditioned predictor into a reusable variational Hamiltonian framework, it addresses the twin challenges of accuracy near phase transitions and the scaling of sampling costs. The authors emphasize that this approach is broadly applicable to various Hamiltonian families (lattice fermions, spins, molecules) and provides a general route toward efficient, foundation-model-based quantum many-body simulation without the prohibitive cost of repeated full optimizations or sampling for every target parameter.

Technical Summary: Foundation Neural Effective Hamiltonian for Strongly Correlated Quantum Materials

More like this