Fragment-Constrained Charge Equilibration for Charge-Aware Machine Learning Potentials at Electrochemical Interfaces

This paper introduces Soft-FQEq, a differentiable fragment-constrained charge equilibration solver integrated into machine learning potentials that enables reactive simulations of electrochemical interfaces by preserving distinct electrode and electrolyte charge distributions and recovering the essential interfacial electrochemical potential gradient, which global charge equilibration methods fail to capture.

The Gist

The Big Picture: Simulating a "Chemical Border"

Imagine you are trying to simulate a busy border crossing between two countries: a solid metal electrode (like a wall) and a liquid electrolyte (like a river of water and salt).

In the real world, this border is special. The metal wall has a specific electrical "mood" (potential), and the water on the other side has a different one. There is a distinct "gradient" or slope of electricity right at the border where they meet. This slope is what drives chemical reactions, like splitting water to make hydrogen fuel.

To simulate this on a computer, scientists use "Machine Learning Potentials" (MLIPs). Think of these as super-smart calculators that predict how atoms move and interact. However, to get the physics right, these calculators need to know how electric charge moves around.

The Problem: The "One-Size-Fits-All" Mistake

The paper explains that the current best way these calculators handle charge is called Global Charge Equilibration (QEq).

The Analogy: Imagine a large party where everyone is holding a balloon. The rule of Global QEq is that everyone must instantly agree on the exact same pressure inside their balloons. If one person's balloon gets a little too full, they instantly share air with everyone else until every single balloon in the room has the exact same pressure.

Why this fails at a border:
In our electrochemical border, the metal wall and the water river are like two different countries. They should have different electrical pressures. But the "Global QEq" rule forces them to equalize instantly.

• The Result: The computer thinks the metal and the water are the same. The "slope" or gradient at the border disappears. The simulation loses the very thing that makes the border interesting. It's like trying to simulate a waterfall by forcing the water at the top and the water at the bottom to be at the exact same height.

The Old Fix: Rigid Topology

Scientists have tried to fix this before using "Per-Fragment" methods.
The Analogy: Instead of letting everyone share air, you put people in separate rooms (fragments). The metal wall is in Room A, and the water molecules are in Room B. They can equalize pressure inside their own room, but not between rooms.

The Catch: This only works if the rooms are fixed. If a water molecule breaks apart or a new bond forms (reactive chemistry), the "room" definition breaks. The computer gets confused because the map of who belongs in which room suddenly changes. It's like trying to use a rigid floor plan for a building where the walls are constantly melting and reforming.

The New Solution: "Soft-FQEq"

This paper introduces a new method called Soft-FQEq (Soft Fragment-Constrained Charge Equilibration).

The Analogy: Instead of rigid walls, imagine the rooms are made of smart, stretchy fog.

1. Dynamic Membership: The computer doesn't need a pre-drawn map. It looks at the atoms and asks, "Are you bonded?" If two atoms are close, the fog is thick between them (they are in the same room). If they are far apart, the fog is thin. If a bond is breaking, the fog just gets thinner gradually.
2. Differentiable Math: Because the "fog" is smooth and mathematically flexible, the computer can handle bonds breaking and forming without crashing. The "rooms" (fragments) change shape and size automatically as the atoms move.
3. The Result: The metal wall stays in its own "foggy room," and the water stays in its own. They can maintain their own electrical pressure (chemical potential) while still talking to each other. This allows the "slope" or gradient at the border to exist naturally.

How They Tested It

The researchers trained this new system on a specific setup: an Iridium Oxide (IrO2) wall with water and salt ions.

1. The Test: They ran the simulation with their new "Soft-FQEq" method.

   • Result: They saw a clear "slope" of electrical potential from the metal wall down to the water. The metal had one value, the water had another, and there was a smooth transition in between. This is exactly what physics predicts should happen.

2. The Control: They took the exact same trained computer brain but swapped out the "Soft-FQEq" solver for the old "Global QEq" solver.

   • Result: The slope vanished. The electrical potential became flat and uniform across the whole system.

The Conclusion: This proved that the "slope" wasn't a lucky accident of the training data. It was a direct result of the new "Soft-FQEq" architecture. The old method physically cannot create that slope, no matter how well you train it.

Why This Matters (According to the Paper)

This isn't just about making better numbers; it's about fixing the fundamental math.

• Reactive Chemistry: Because the "fog" (fragment identification) is flexible, this method can handle chemical reactions where bonds break and form, which rigid methods cannot do.
• Realistic Interfaces: It allows scientists to finally simulate electrochemical interfaces (like batteries or fuel cells) where the metal and liquid have distinct electrical personalities, without forcing them to be identical.

In short, the paper built a new "mathematical lens" that lets computers see the electrical differences between a metal and a liquid, even when they are reacting and changing shape, which previous methods were too rigid to see.

Technical Summary

1. Problem Statement

Predictive atomistic modeling of electrochemical interfaces (e.g., electrode-electrolyte systems) requires machine learning interatomic potentials (MLIPs) that can simultaneously handle:

• Reactive bond breaking/forming.
• Long-range electrostatics.
• Spatially varying charge distributions reflecting the distinct electronic nature of the electrode versus the electrolyte.

The Core Limitation: Existing charge-aware MLIPs typically rely on Global Charge Equilibration (QEq). In global QEq, a single system-wide Lagrange multiplier ($\mu_{global}$) enforces total charge conservation. This forces the entire system (electrode, solvent, and ions) to equilibrate to a single electrochemical potential.

• Consequence: This structural constraint prevents the formation of the necessary electrochemical potential gradient across the interface (the electric double layer).
• Pathology: It leads to spurious long-range charge transfer between electronically disconnected regions (e.g., charge flowing from the bulk electrolyte to the electrode without a physical pathway), regardless of how well the neural network parameters are trained.
• Existing Gap: Classical molecular dynamics uses "per-fragment" equilibration to solve this, but it relies on predefined, hard-wired molecular topology, making it incompatible with reactive systems where bonds break and form.

2. Methodology: Soft-FQEq

The authors propose Soft-FQEq (Soft Fragment-Constrained Charge Equilibration), a differentiable solver layer that restores fragment-resolved charge conservation in reactive MLIPs without relying on fixed topology.

A. Differentiable Fragment Identification

Instead of hard assignment of atoms to fragments, the method learns fragment membership as a continuous function of geometry:

1. Soft Bond Connectivity: A Multi-Layer Perceptron (MLP) predicts a "soft bond connectivity" ($\kappa_{ij} \in [0, 1]$) for every atom pair based on interatomic distances and learned features. Values near 1 indicate a bond; values near 0 indicate no bond.
2. Resolvent Construction: A graph Laplacian resolvent is constructed from these soft bond weights. This matrix is smoothed and sharpened using an effective resistance metric to generate a soft fragment membership matrix ($\tilde{C}$).
   • $\tilde{C}$ is $C^\infty$ smooth, allowing atoms to migrate between fragments continuously during bond-breaking/forming events without discontinuities in the potential energy surface (PES).
   • It satisfies global connectivity (e.g., treating a conductive electrode slab as a single fragment) and local uniformity.

B. The Solver Layer (Augmented-Lagrangian Uzawa)

The charge equilibration problem is reformulated with constraints per fragment rather than per system:

• Objective: Minimize electrostatic energy $E_{elec} = \chi^T q + \frac{1}{2}q^T A q$ subject to global charge conservation ($1^T q = Q_{total}$) and per-fragment charge conservation ($\tilde{C}^T q = Q^*$).
• Algorithm: An Augmented-Lagrangian Uzawa iteration is used to solve the resulting Karush-Kuhn-Tucker (KKT) system.
  • It introduces a fragment-level Lagrange multiplier vector ($\mu_{frag}$).
  • The solver is unrolled for a fixed number of steps ($K$) to allow backpropagation of gradients through the charge solution to the upstream neural network heads.
• Outputs: The solver returns equilibrated charges ($q$) and per-fragment chemical potentials ($\mu_{frag}$).

C. Architecture and Training

• Base Network: Implemented on the HIP-NN (Hierarchically Interacting Particle Neural Network) framework.
• Readout Heads: The shared feature network feeds four scalar MLP heads:
  1. Per-atom short-range energy ($E_{short}$).
  2. Per-atom electronegativity ($\chi$).
  3. Per-atom source charge ($q_c$).
  4. Per-pair bond connectivity correction ($\Delta \log \kappa$).
• Training Strategy:
  • Trained on DFT energies, forces, and DDEC6 charges (chosen for their faithful representation of electrostatic potentials).
  • Auxiliary Losses: Crucial for breaking the "compensatory symmetry" where the solver fixes errors in the neural network. Specific losses supervise $\chi$ and $q_c$ directly to ensure physical meaning, preventing the solver from masking poor predictions with Lagrange multipliers.
  • Gauge Fixing: Electronegativity is centered to zero mean to remove gauge freedom.

3. Key Contributions

1. Soft-FQEq Solver: A novel, differentiable solver layer that enables per-fragment charge equilibration in reactive MLIPs by making fragment identification a continuous function of atomic geometry.
2. Structural Resolution of the Double Layer: Demonstrates that the electrochemical potential gradient is a structural requirement of the solver, not just a result of better training data or parameterization.
3. Architecture Independence: The solver couples to the MLIP only through scalar readouts, making it applicable to various feature networks (e.g., MACE, NequIP).
4. Ablation Study: Provides a rigorous proof that global QEq fundamentally cannot support interfacial gradients, while fragment-constrained QEq can, using the exact same trained weights.

4. Results

The model was trained and validated on IrO$_2$/H$_2$O/Na$^+$/ClO$_4^-$ interfaces.

• Charge Accuracy: The model reproduces DDEC6 reference charges with a Mean Absolute Error (MAE) of $\sim 0.05$ e.
• Electrochemical Potential Gradient ($\mu_{phys}$):
  • Soft-FQEq: The model successfully recovers a distinct spatial profile: a flat plateau in the electrode, a sharp gradient at the interface, and a distinct plateau in the bulk electrolyte. This represents the physical electric double layer.
  • Global QEq (Ablation): When the trained weights are used with a global QEq solver (replacing the fragment-constrained one), the interfacial gradient collapses completely to a uniform value across the entire system ($< 10^{-6}$ eV variation). This proves the gradient is impossible to sustain under global QEq.
• Response to Charging: The model correctly predicts a monotonic shift in the electrode's electrochemical potential as the surface charge density is varied (capacitive response), while the bulk electrolyte remains relatively stable.
• Fragment Identification: The soft bond connectivity correctly identifies the electrode as a single fragment, intact water molecules as separate fragments, and individual ions, even during thermal fluctuations.

5. Significance

• Enabling Reactive Electrochemistry: This work bridges the gap between classical constant-potential methods (which handle electrostatics but not reactivity) and reactive MLIPs (which handle reactivity but fail at electrostatics). It allows for the simulation of reactive electrochemical double layers with near-DFT accuracy.
• Fundamental Insight: It establishes that the failure of global QEq to model interfaces is a mathematical structural flaw (the single-$\mu$ constraint), not a training limitation.
• Future Applications: The framework paves the way for:
  • Constant-potential molecular dynamics (via extractable electrode potentials).
  • Simulations of complex systems like ionic liquids, molten salts, and supported metal clusters where electronically distinct subsystems exist without covalent bonding.
  • Integration with Grand Canonical Monte Carlo (GCMC) for true constant-potential sampling.

In summary, the paper introduces a mathematically rigorous and differentiable method to enforce local charge conservation in reactive MLIPs, successfully resolving the long-standing inability of charge-aware potentials to model the electrochemical potential gradient at interfaces.