More converged, less accurate? Reassessing standard choices for ab initio water using machine learning potentials

This study demonstrates that using highly converged electronic structure calculations is essential for accurately modeling liquid water and ice, revealing that commonly used pragmatic settings often yield misleading agreement with experiments while more rigorous approaches significantly improve predictive accuracy.

The Gist

Imagine you are trying to build the perfect digital simulation of water. You want it to look, flow, and freeze exactly like the real stuff in your glass. To do this, scientists use powerful computer models based on the laws of quantum physics. But here's the catch: these models are like high-end cameras. If you don't adjust the settings correctly, the photo might look "good enough" by accident, even if the camera is actually misfocusing.

This paper, titled "More converged, less accurate?", is a detective story about how scientists accidentally found that their "good enough" settings were actually hiding the truth.

Here is the breakdown of their discovery, using some everyday analogies:

1. The "Good Enough" Trap

For years, scientists have used a specific recipe (called revPBE0-D3) to simulate water. It was famous because it matched real-world experiments perfectly. It was like a chef who always made a delicious soup that tasted exactly like the restaurant's signature dish.

However, this chef was using a "quick and dirty" method. They were using a smaller, cheaper set of ingredients (a triple-ζ basis set) and a simplified map of the atoms (using pseudopotentials, which ignore the inner core of atoms to save time).

The Analogy: Imagine trying to draw a portrait of a friend. You use a cheap sketchbook and a blunt pencil. Surprisingly, the drawing looks just like your friend! You think, "Wow, my sketching skills are amazing." But in reality, the cheap pencil and paper just happened to blur the lines in a way that accidentally made the nose look right.

2. The "High-Definition" Test

The authors of this paper decided to stop using the "quick and dirty" method. They switched to the "High-Definition" mode:

• Better Ingredients: They used a massive, ultra-detailed set of ingredients (a quadruple-ζ basis set).
• Real Maps: They stopped using the simplified maps and looked at every single electron (using all-electron potentials).
• Tighter Rules: They made the computer calculate with extreme precision, leaving no room for error.

The Shocking Result: When they used these "perfect" settings with the same famous recipe (revPBE0-D3), the simulation stopped looking like real water. The water became too structured, the molecules moved too slowly, and the density was wrong.

The Lesson: The previous "perfect" match wasn't because the recipe was great; it was because the errors in the cheap settings canceled out the errors in the recipe. It was a "happy accident." When they fixed the settings, the accident disappeared, and the recipe looked flawed.

3. The New Champion

So, if the famous recipe is flawed, what works?
The scientists tried a different recipe called řB97X-rV. When they cooked this one with the "High-Definition" settings, it actually tasted like the real thing! It matched the experimental data much better than the old "accidental" winner.

4. The "MP2" Disaster

They also tried a method called MP2, which is often considered the "gold standard" for accuracy in chemistry. But, they used a small, cheap ingredient list for it.
The Result: The simulation was a total disaster. The water molecules clumped together too tightly, like a frozen block of ice that wouldn't melt, and they barely moved.
The Lesson: Even a "gold standard" method fails if you don't give it enough computational power (a big enough basis set) to do its job.

5. The "Egg Box" Effect

One of the most interesting technical discoveries in the paper is something called the "Egg Box Effect."
Imagine you are placing eggs in a carton. If the carton is slightly crooked, the eggs sit at slightly different heights. In computer simulations, if the grid the computer uses to calculate energy is slightly misaligned, the energy of the water molecules changes just because of where they are sitting on the grid.

The authors found that their old, cheap settings had a lot of this "wobble" (noise). The machine learning models they trained learned to ignore this noise, which accidentally helped them match real-world data. But when they fixed the "wobble" (using a better calculation method called GAPW), the noise disappeared, and the models had to face the raw truth of the physics.

The Big Takeaway

This paper teaches us a vital lesson for the future of science:

"More converged, less accurate" is a warning.

Just because a simulation matches real-world data doesn't mean the underlying theory is perfect. It might just mean the errors in the calculation settings accidentally canceled out the errors in the theory.

To build truly reliable models for water (and other complex systems), we can't just rely on "good enough" shortcuts. We need to use fully converged, high-precision calculations to train our AI. Only then can we be sure we aren't just getting lucky with a blurry sketch, but actually understanding the true nature of water.

In short: The authors turned up the resolution on their computer models, and suddenly, the "perfect" water they thought they had was revealed to be a mirage. The real winner was a different recipe that only shines when you give it the best possible tools.

Technical Summary

1. Problem Statement

Accurately simulating liquid water remains a central challenge in computational chemistry. While Density Functional Theory (DFT) is the standard tool, the choice of exchange-correlation functionals, basis sets, and convergence settings significantly impacts predicted structural and dynamical properties.

• The Paradox: The hybrid functional revPBE0-D3 (specifically with a triple-ζ basis set and pseudopotentials) has historically shown impressive agreement with experimental data for water properties (density, diffusion, radial distribution functions).
• The Issue: It is unclear whether this agreement stems from the intrinsic accuracy of the functional or a fortuitous cancellation of errors arising from insufficient convergence (e.g., basis set incompleteness, pseudopotential approximations, and the "egg box effect" in plane-wave calculations).
• The Opportunity: Machine Learning Potentials (MLPs) allow for long molecular dynamics (MD) simulations at a computational cost far lower than ab initio methods. This enables researchers to move beyond the "cost-accuracy trade-off" and perform fully converged reference calculations to determine the true performance of electronic structure methods.

2. Methodology

The authors employed a rigorous workflow combining high-level electronic structure calculations with Committee Neural Network Potentials (C-NNPs) to decouple methodological errors from physical predictions.

• Reference Electronic Structure Setups: Four distinct setups were used to generate training data:

  1. Baseline: revPBE0-D3/TZV2P/GTH (Standard triple-ζ basis, GTH pseudopotentials, Gaussian and Plane Waves (GPW) method).
  2. Highly Converged revPBE0-D3: revPBE0-D3/def2-QZVP/AE (Quadruple-ζ basis, All-Electron (AE) potentials, Gaussian and Augmented Plane Waves (GAPW) method, tighter SCF convergence).
  3. Alternative Functional: řB97X-rV/def2-QZVP/AE (Range-separated hybrid functional with the same high-convergence settings as #2).
  4. Post-HF Method: MP2/cc-TZ/GTH (Second-order Møller–Plesset perturbation theory with triple-ζ basis and pseudopotentials).

• Machine Learning Model:

  • Four Committee Neural Network Potentials (C-NNPs) were trained using the N2P2 package (Behler–Parrinello architecture).
  • Training data included 964 structures from liquid water, ice (Ih, VIII), and water slabs, sampled via Query by Committee (QbC) to ensure diversity.
  • The dataset was extended with NpT (constant pressure) simulations to improve density predictions.

• Simulation Protocol:

  • Classical MD and Path-Integral MD (PIMD) to include Nuclear Quantum Effects (NQEs) were performed using the i-PI package.
  • Observables Calculated: Oxygen-Oxygen Radial Distribution Function (RDF), pressure-density curves, diffusion coefficients, hydrogen bond lifetimes, and covalent bond lengths.
  • Validation: Models were validated against independent test sets, with a specific focus on "net forces" (sum of forces in the system) to detect numerical noise (egg box effect).

3. Key Contributions

• Identification of the "Egg Box" Effect in Water Datasets: The authors demonstrated that standard GPW calculations with lower cutoffs suffer from the "egg box effect" (grid-dependent energy oscillations), leading to non-zero net forces in reference datasets. Switching to the GAPW method and increasing plane-wave cutoffs significantly reduced these numerical artifacts.
• Re-evaluation of revPBE0-D3: They challenged the consensus that revPBE0-D3 is inherently superior for water, showing its agreement with experiment is likely an artifact of error cancellation between the functional and the incomplete basis set/pseudopotentials.
• Systematic Benchmarking: Provided a comprehensive comparison of DFT and MP2 methods for water under both standard and fully converged conditions, isolating the effects of basis sets, pseudopotentials, and functionals.

4. Key Results

A. Model Validation & Numerical Artifacts

• Net Forces: The standard revPBE0-D3/TZV2P/GTH dataset exhibited an average net force of 2.14 meV/Å/atom, indicating significant numerical noise. The highly converged setups (def2-QZVP/AE) reduced this to <0.1 meV/Å/atom.
• Impact on Errors: Models trained on the "noisy" standard setup showed higher test-set errors when evaluated against the "clean" converged reference, not because the model was worse, but because the reference data contained unlearnable noise (egg box effect).

B. Structural Properties (RDF)

• revPBE0-D3: The standard setup (TZV2P/GTH) reproduced experimental RDFs almost perfectly. However, the highly converged setup (def2-QZVP/AE) resulted in a flatter first peak and a less structured RDF, deviating significantly from experiment. This suggests the standard setup's accuracy was accidental.
• řB97X-rV: The highly converged řB97X-rV setup showed better agreement with experiment than the converged revPBE0-D3, though still not perfect.
• MP2: The MP2/cc-TZ/GTH setup produced a severely over-structured RDF (peaks too high, valleys too deep), indicating the triple-ζ basis set is insufficient for MP2 in bulk water.

C. Thermodynamic Properties (Density)

• Density: The standard revPBE0-D3/TZV2P/GTH underestimated water density. The highly converged revPBE0-D3/def2-QZVP/AE significantly improved density predictions, matching the experimental curve closely at low pressures.
• řB97X-rV: Overestimated the density of liquid water across the pressure range.
• MP2: Provided a decent pressure-density curve for liquid water (despite the poor RDF) but overestimated the density of ice Ih.

D. Dynamical Properties (Diffusion & Hydrogen Bonds)

• Diffusion Coefficient:
  • Standard revPBE0-D3 matched experimental diffusion well.
  • Converged revPBE0-D3 showed faster diffusion (less structured water).
  • MP2 severely underestimated diffusion (water was too "stuck"), with a large portion of the calculated diffusion coming from finite-size corrections rather than actual dynamics.
• Hydrogen Bond Lifetimes:
  • MP2 exhibited extremely long hydrogen bond lifetimes (~20 ps), consistent with its over-structured nature.
  • Converged DFT setups showed shorter lifetimes than the standard setup, correlating with the increased diffusion.

E. Role of NQEs

• Including NQEs generally increased diffusion and slightly altered bond lengths.
• For the converged revPBE0-D3, NQEs had a negligible effect on density, contrasting with previous studies using the standard setup.

5. Significance and Conclusion

• Convergence is Critical: The study proves that "standard" computational settings (triple-ζ basis, pseudopotentials) are insufficient to reveal the true accuracy of electronic structure methods for water. The apparent success of revPBE0-D3 is largely due to the cancellation of errors between the functional, the basis set, and the pseudopotential.
• MLPs as a Tool for Convergence: MLPs enable the routine use of highly converged, expensive reference calculations (like All-Electron GAPW or high-level MP2) for MD simulations, which were previously computationally prohibitive.
• Methodological Recommendations:
  • Researchers should use GAPW methods and tighter convergence criteria to eliminate the "egg box effect" and ensure reliable reference data.
  • MP2 requires basis sets larger than triple-ζ (or fragment-based approaches) to be reliable for bulk water.
  • řB97X-rV emerges as a more robust functional than revPBE0-D3 when used with fully converged settings, despite the latter's historical popularity.

In summary, the paper argues that more converged calculations often yield results that are less aligned with experiment when using standard functionals, not because the functionals are wrong, but because the "standard" agreement was an illusion created by numerical errors. This necessitates a re-evaluation of the "gold standard" setups for aqueous simulations.