Fast Evaluation of Unbiased Atomic Forces in ab initio Variational Monte Carlo via the Lagrangian Technique

This paper introduces a Lagrangian-based technique that replaces the computationally expensive requirement of $6N$ DFT calculations with a single coupled-perturbed Kohn-Sham calculation to efficiently generate unbiased atomic forces in ab initio Variational Monte Carlo, thereby improving their consistency with potential energy surfaces and accuracy relative to CCSD(T) benchmarks.

The Gist

The Big Picture: The "Perfect Map" Problem

Imagine you are trying to build a GPS navigation system for a car (the car is an atom, and the road is the molecule it's part of). To make the GPS work perfectly, you need two things:

1. The Map (Potential Energy Surface): A perfect drawing of the terrain showing hills and valleys.
2. The Steering Wheel (Atomic Forces): The actual push and pull you feel when you turn the wheel, telling the car exactly which way to go.

In the world of quantum chemistry, scientists use a super-accurate method called Quantum Monte Carlo (QMC) to draw these maps. It's like using a million tiny drones to survey a landscape; it's incredibly precise but very slow.

The Problem:
For a long time, while QMC could draw the map perfectly, it was terrible at calculating the steering wheel (the forces).

• The "Biased" Force: The old way of calculating forces was like guessing the steering direction based on a blurry photo. It usually worked, but sometimes it pushed the car in the wrong direction, leading to a crash (inaccurate simulations).
• The "Unbiased" Force: This is the perfect steering wheel that matches the map exactly.

The Old Fix: The "Brute Force" Method

A few years ago, the authors of this paper found a way to get the perfect steering wheel. But it was incredibly expensive.

The Analogy:
Imagine you want to know how the steering wheel reacts if you move the car's front bumper just a tiny bit to the left, right, up, or down.

• To get the perfect answer using the old method, you had to build a brand new, perfect map for every single tiny movement of every single atom.
• If your car had 10 atoms, you had to build 60 new maps (6 directions × 10 atoms).
• If your car had 1,000 atoms, you had to build 6,000 maps.

This was like trying to survey a whole country by walking every single inch of every street. It was too slow to be useful for big systems or for training AI robots to drive cars.

The New Solution: The "Lagrangian Shortcut"

This paper introduces a clever mathematical trick called the Lagrangian Technique.

The Analogy:
Instead of walking every inch of the street to see how the road changes, imagine you have a smart surveyor who understands the rules of how the road is built.

• You give the surveyor the original map.
• You ask: "If I nudge this atom here, how does the whole road react?"
• The surveyor uses a set of rules (mathematical equations) to calculate the answer instantly, without needing to rebuild the whole map 6,000 times.

In the paper, this "smart surveyor" is a single calculation called a Coupled-Perturbed Kohn-Sham (CPKS) calculation.

• Old Way: 6,000 separate calculations (Slow, expensive, impossible for big systems).
• New Way: 1 single calculation (Fast, cheap, scalable).

Why Does This Matter? (The "Machine Learning" Connection)

Why do we care about perfect steering wheels? Because we are teaching Artificial Intelligence (AI) to predict how molecules behave.

• The AI Student: To learn how to drive, the AI needs a teacher. The teacher provides data: "Here is the map, and here is the force you should feel."
• The Problem: If the teacher gives the AI a blurry, biased steering wheel, the AI learns bad habits. It might think a hill is a valley.
• The Result: The AI builds a bad model. It might predict that a drug molecule works when it doesn't, or that a new battery material is stable when it explodes.

By using this new "Lagrangian Shortcut," the authors can now generate perfect, unbiased data for the AI to learn from, even for very large molecules. This means we can train AI to discover new medicines and materials much faster and more accurately.

The "Gold Standard" Check

To prove their new method works, the authors tested it on three molecules (Ethanol, Malonaldehyde, and Benzene). They compared their results against the "Gold Standard" of chemistry (a method called CCSD(T)), which is the most accurate method we have, but also the slowest.

• The Old Biased Force: Was like a student who got a C- on the test. It was okay, but not good enough for serious work.
• The New Unbiased Force: Got an A. It matched the Gold Standard much better.

The Surprise:
They found that for some molecules, the new QMC forces were actually better than some popular, faster methods (like certain types of Density Functional Theory). This suggests that QMC, once we fix the force calculation, is a powerhouse for the future of material science.

Summary in One Sentence

The authors found a mathematical "shortcut" that allows super-accurate quantum computers to calculate the forces on atoms without needing to do thousands of extra, slow calculations, making it possible to train AI to design new materials and drugs with unprecedented accuracy.

Technical Summary

1. Problem Statement

Ab initio Quantum Monte Carlo (QMC) methods, particularly Variational Monte Carlo (VMC), are state-of-the-art for solving the many-body Schrödinger equation with high accuracy and parallel scalability. However, their application to large systems and machine learning potential (MLP) training is hindered by the difficulty of computing unbiased atomic forces (forces consistent with the potential energy surface, PES).

• The Bias Issue: In VMC, atomic forces are typically calculated using the Hellmann-Feynman (HF) and Pulay terms. If the wavefunction is not fully optimized (a common scenario where the Slater determinant is frozen from a Density Functional Theory (DFT) calculation and only the Jastrow factor is optimized), a "non-variational" (NV) term arises. Neglecting this NV term introduces a Self-Consistency Error (SCE), leading to biased forces that do not match the true PES gradient.
• The Computational Bottleneck: A previous method by Nakano et al. (2024) successfully eliminated this bias by calculating the NV term using a hybrid VMC-DFT approach. However, this required $6N$ additional DFT calculations (where $N$ is the number of nuclei) to compute finite differences for the wavefunction response. This scaling ($O(N^4)$ effectively) makes the method computationally prohibitive for large systems or extensive datasets required for MLP training.

2. Methodology: The Lagrangian Technique

The authors propose a new method to compute unbiased forces by extending the Lagrangian technique (standard in quantum chemistry for coupled-perturbed calculations) to the VMC framework.

• Core Concept: Instead of using finite differences (FDM) to evaluate the wavefunction response ($\partial p_i / \partial R_\alpha$), the method constructs a Lagrangian function ($L_{VMC}$) that incorporates the VMC energy and the constraints of the underlying SCF (DFT) reference calculation (orbital orthonormality and stationarity).
• Stationarity Condition: By making the Lagrangian stationary with respect to all parameters (Jastrow parameters, MO coefficients, and Lagrange multipliers), the authors derive a fully variational expression for the force.
• Z-Vector Equations: The method replaces the $6N$ DFT calculations with a single-shot Coupled-Perturbed Kohn-Sham (CPKS) or Coupled-Perturbed Hartree-Fock (CPHF) calculation.
  • The response of the orbitals to the Jastrow correlation is determined by solving a linear system of equations known as the Z-vector equations ($AZ = -B$).
  • The vector $B$ contains the derivatives of the VMC energy with respect to the MO coefficients (computed via VMC sampling).
  • The matrix $A$ is derived from the SCF stationary conditions and is independent of the VMC noise.
• Implementation: The method was implemented by interfacing the CP2K software (for DFT and CPKS) with TurboRVB (for VMC), utilizing the TREX-IO library for standardized data exchange.

3. Key Contributions

1. Algorithmic Innovation: Successfully extended the Lagrangian technique to VMC, enabling the calculation of unbiased forces with a single CPKS/CPHF step instead of $6N$ DFT steps.
2. Scalability Improvement: Reduced the computational scaling for the force correction from effectively $O(N^4)$ (due to $6N$ DFT calls) to $O(N^3)$ (scaling of a single CPKS calculation), making unbiased VMC forces feasible for larger systems.
3. Validation: Demonstrated that the Lagrangian-derived forces are numerically identical to those obtained via the finite-difference method (FDM) and consistent with the numerical derivatives of the potential energy surface (PES).
4. Accuracy Benchmarking: Showed that correcting the NV term not only removes bias (consistency with PES) but also improves the accuracy of the forces relative to high-level benchmarks.

4. Results and Benchmarks

The authors validated the method on three systems: the Cl$_2$ molecule, cubic Boron Nitride (c-BN) crystal, and an Acetamide (AcNH$_2$) dimer.

• Consistency: For all test cases, the forces and pressures calculated via the Lagrangian technique (LR) matched the reference values derived from numerical derivatives of the PES and the previous FDM-based NV correction.
• Performance: Timing benchmarks on c-BN systems showed that the single-shot LR calculation is significantly faster than the $6N$ DFT approach. For a system with 4,096 atoms, the LR calculation took ~134 minutes, whereas the $6N$ DFT approach would have taken over 27,000 hours (extrapolated).
• Accuracy on rMD17 Dataset: The method was tested on ethanol, malonaldehyde, and benzene from the rMD17 benchmark set, comparing VMC forces against CCSD(T) (Coupled-Cluster Singles and Doubles with perturbative Triples) as the gold standard.
  • Bias Reduction: Unbiased VMC forces showed significantly lower Root-Mean-Square Errors (RMSE) compared to biased VMC forces.
    • Ethanol: RMSE dropped from 2.92 to 2.55 kcal/mol/Å.
    • Malonaldehyde: RMSE dropped from 9.56 to 7.00 kcal/mol/Å.
    • Benzene: RMSE dropped from 3.19 to 2.18 kcal/mol/Å.
  • Comparison with Wavefunction Theory:
    • Unbiased VMC forces achieved accuracy comparable to or better than fully optimized VMC wavefunctions (where the determinant is also optimized), proving that correcting the NV term is a highly efficient alternative to full wavefunction optimization.
    • For ethanol and benzene, unbiased VMC forces were competitive with hybrid/meta-GGA DFT functionals (e.g., $\omega$B97X-D3BJ).
    • For malonaldehyde, both VMC and CCSD(T) showed large discrepancies with other methods. The authors attribute this to the multi-reference character of malonaldehyde (indicated by high T1 diagnostics), suggesting that single-reference CCSD(T) may struggle here, and VMC forces might actually be more physically reliable in this specific regime.

5. Significance and Future Outlook

• Machine Learning Potentials: This work removes a major bottleneck in generating high-quality training data for Machine Learning Interatomic Potentials (MLPs). Unbiased VMC forces can now be generated efficiently for large datasets, providing a more accurate alternative to DFT for training MLPs, especially for systems where DFT fails (e.g., strong correlation, dispersion).
• General Applicability: While demonstrated with Gaussian Type Orbitals (GTOs) and Jastrow-Slater determinants, the Lagrangian formalism is versatile and can be adapted to other ansatzes (e.g., multi-determinant) and basis sets (e.g., plane waves).
• Future Work: The authors note that while the method is applicable to Diffusion Monte Carlo (DMC) in principle, a full DMC implementation is left for future work. Additionally, further investigation into multi-reference systems (like malonaldehyde) using multi-determinant VMC ansatzes is proposed.

In summary, this paper provides a critical methodological advancement that makes unbiased, high-accuracy atomic forces computationally affordable in VMC, bridging the gap between high-fidelity quantum chemistry and the data requirements of modern machine learning potentials.