mole-lambda-coupled-cluster-response

name	mole-lambda-coupled-cluster-response
description	MōLe-Λ methodology for learning coupled-cluster response states. Extends Molecular Orbital Learning (MōLe) to predict full CCSD response state by jointly learning T and Λ amplitudes from localized Hartree-Fock orbitals. Provides CC-quality energies, forces, dipoles, polarizabilities, electron density at ML speed. ICML 2026 AI4Physics. Activation: coupled-cluster, CCSD response, molecular orbital learning, quantum chemistry surrogate, Λ-amplitudes, equivariant quantum chemistry, wavefunction learning
metadata	{"arxiv_id":"2605.29622","published":"2026-05-28","authors":"Andreas Burger, Luca Thiede, Abdulrahman Aldossary, Jorge A. Campos-Gonzalez-Angulo, Alex Zook, Jérôme Florian Gonthier, Alán Aspuru-Guzik","tags":["quantum-chemistry","coupled-cluster","equivariant-ml","molecular-orbitals","surrogate-model","response-theory"]}

MōLe-Λ: Learning the Coupled-Cluster Response State

Core Insight

Coupled-cluster (CC) theory is the gold standard of quantum chemistry, but its O(N⁶) computational cost limits routine access to accurate energies, forces, and response properties. MōLe-Λ extends the Molecular Orbital Learning (MōLe) framework to learn the full CCSD response state — not just the wavefunction (T amplitudes) but also the response (Λ amplitudes) — enabling prediction of all CC-quality observables at ML inference speed.

Key Architecture Components

1. Joint T + Λ Amplitude Learning

The coupled-cluster equations have two sides:

• Right-hand (T) amplitudes: Determine the correlated wavefunction |Ψ⟩ = e^T |Φ₀⟩
• Left-hand (Λ) amplitudes: Required for computing response properties (forces, dipoles, polarizabilities)

MōLe-Λ learns both from localized Hartree-Fock molecular orbitals:

(T, Λ) = MōLe-Λ(orbital_features)

2. Symmetry-Consistent Readout Heads

class MoLeLambdaArchitecture:
    """
    Extends MōLe with Λ and T readout heads that mirror
    the symmetry constraints of the CCSD equations.
    """
    def __init__(self):
        self.encoder = EquivariantOrbitalEncoder()  # Original MōLe encoder
        self.T_head = OddSignEquivariantDecoder()    # T amplitudes (antisymmetric)
        self.Lambda_head = OddSignEquivariantDecoder()  # Λ amplitudes (antisymmetric)

    def forward(self, localized_orbitals):
        features = self.encoder(localized_orbitals)
        T = self.T_head(features)
        Lambda = self.Lambda_head(features)
        return T, Lambda

3. Preserved Invariance Properties

MōLe-Λ preserves the key physical invariances:

• Equivariant orbital encoder: Respects rotational/translational symmetry
• Odd sign-equivariant decoding: Correct behavior under orbital sign flips
• Locality: Exploits spatial decay of electron correlation
• Size-extensivity: Energy scales correctly with system size

Observable Prediction

MōLe-Λ recovers all CCSD-quality observables:

Observable	Requires	MōLe-Λ Support
Energy	T only	✓
Forces (gradients)	T + Λ	✓
Dipole moments	T + Λ	✓
Quadrupole moments	T + Λ	✓
Polarizabilities	T + Λ	✓
Electron density	T + Λ	✓
Pair density (2-electron)	T + Λ	✓

Speed Advantage

Full CCSD:     O(N⁶) — expensive, scales poorly
MōLe-Λ:        O(N) inference — after training, near-constant cost
Speedup:       ~100-1000x for large molecules

Workflow

1. Prepare localized Hartree-Fock molecular orbitals
2. Encode orbitals with equivariant encoder
3. Decode T and Λ amplitudes via symmetry-consistent readout heads
4. Compute observables from (T, Λ) pair using standard CC formulas
5. Validate against reference CCSD calculation

Practical Usage

# Pseudocode for MōLe-Λ inference
import torch

def compute_cc_properties(molecule, model):
    """Compute CC-quality properties using MōLe-Λ."""
    # Step 1: Get localized HF orbitals
    hf_orbitals = get_localized_orbitals(molecule)

    # Step 2: Encode + decode
    T, Lambda = model(hf_orbitals)

    # Step 3: Compute observables
    energy = cc_energy(T)
    forces = cc_forces(T, Lambda)
    dipole = cc_dipole(T, Lambda)
    polarizability = cc_polarizability(T, Lambda)
    density = cc_electron_density(T, Lambda)
    pair_density = cc_pair_density(T, Lambda)

    return {
        'energy': energy,
        'forces': forces,
        'dipole': dipole,
        'polarizability': polarizability,
        'density': density,
        'pair_density': pair_density,
    }

When to Apply

• Surrogate modeling: Replace expensive CCSD calculations with ML inference
• High-throughput screening: Scan thousands of molecules at CC quality
• Response property prediction: When forces, dipoles, or polarizabilities are needed
• Molecular dynamics: CC-quality forces at ML speed enables ab initio MD

Key Findings

1. Joint learning works: T and Λ can be learned simultaneously from localized orbitals
2. Symmetry matters: Enforcing CCSD symmetry constraints in readout heads is critical
3. Property expansion: Extending from energies-only to full response properties
4. Speed without sacrifice: Near-CCSD accuracy at ML inference speed

Pitfalls

• Training data: Requires pre-computed CCSD reference data (T, Λ pairs) for training
• Transferability: Model trained on specific chemical space may not generalize
• Orbital localization: Results depend on quality of localized orbital construction
• Extrapolation: ML surrogate may fail on molecules outside training distribution

References

• Paper: "MōLe-Λ: Learning the Coupled-Cluster Response State for Energies, Gradients, and Properties" (arXiv:2605.29622)
• Conference: ICML 2026 AI4Physics
• Categories: cs.LG, physics.chem-ph
• Date: May 28, 2026