Explicit Electric Potential-Embedded Machine Learning Framework: A Unified Description from Atomic to Electronic Scales

This paper proposes a unified machine learning framework integrating Potential-Embedded MACE and Potential-Embedded Electron Density Prediction to simultaneously and accurately simulate atomic forces and electron density distributions across arbitrary electric potentials, enabling large-scale studies of electrochemical interfaces like the Pt(111)/water system.

The Gist

Imagine you are trying to understand how a battery works or how a chemical reaction happens on a metal surface submerged in water. To do this, scientists need to simulate two things happening at once:

1. The Dance of Atoms: How the water molecules and metal atoms move around (like a crowded dance floor).
2. The Flow of Electricity: How the invisible cloud of electrons shifts and changes to create the electric current (like the electricity flowing through the wires).

Traditionally, doing this is like trying to film a high-speed race with a camera that takes one photo every hour. It's too slow and too expensive to see the whole race.

This paper introduces a new "Smart Simulator" (a Machine Learning Framework) that acts like a super-fast, super-smart camera. It can predict both the dance of the atoms and the flow of electricity instantly, even when you change the voltage (the "push" of the electricity).

Here is how their new system works, broken down into simple parts:

1. The Problem: The "Leaky Bucket" and the "Slow Calculator"

• The Old Way (DFT): The most accurate way to simulate these reactions is called DFT (Density Functional Theory). But it's incredibly slow. It's like trying to calculate the weather for a whole city by measuring the temperature of every single raindrop. You can't do it for a long time or a big area.
• The "Leaky Bucket" Issue: When scientists try to simulate water touching metal, older computer models sometimes make a mistake where water molecules "leak" into the metal in a way that's physically impossible, like water passing through a solid wall.
• The Missing Piece: Existing "fast" AI models could predict how atoms move, but they couldn't tell you what the electrons were doing. It's like having a movie of the dancers but no sound of the music.

2. The Solution: A Unified "Electric Brain"

The authors built a system with three main parts, like a three-step assembly line:

Step A: The Data Factory (Hy DFT)

First, they needed to teach the AI. They built a special software tool called Hy DFT.

• Analogy: Think of this as a high-speed camera that takes perfect snapshots of the battery interface.
• What it does: It simulates the reaction while keeping the electric voltage steady (like holding a constant pressure on a hose). Crucially, it fixes the "leaky bucket" problem so the water stays where it should. It also tags every single snapshot with the exact voltage level, creating a massive library of "What happens at -1 Volt?" vs. "What happens at +1 Volt?"

Step B: The Two-Headed AI (PE-MACE & PE-EDP)

They trained two separate but connected AI brains using the data from Step A. Both brains have a special trick: they treat the Electric Voltage as a basic ingredient, just like they treat "Hydrogen" or "Oxygen" as ingredients.

• Brain 1: PE-MACE (The Motion Tracker)

  • Job: Predicts how atoms move and push against each other.
  • Analogy: Imagine a choreographer who knows exactly how the dancers (atoms) will step, jump, and spin when the music (voltage) changes.
  • Result: It can run simulations for nanoseconds (which is a long time in computer science) instead of picoseconds, revealing how water molecules reorganize themselves when the voltage changes.

• Brain 2: PE-EDP (The Electron Mapper)

  • Job: Predicts the shape and density of the electron cloud.
  • Analogy: Imagine a 3D artist who can instantly paint the invisible "glow" of electricity around the atoms.
  • Result: It tells us exactly where the electrons are, which is vital for understanding how chemical reactions actually happen. Before this, you had to stop the movie and do a slow calculation to see the electrons; now, the AI paints them instantly.

3. The Test Drive: The Platinum-Water Interface

To prove it works, they tested it on a Platinum (Pt) surface with water. This is a classic setup for fuel cells.

• The Discovery: They ran a 4-nanosecond simulation (a huge leap in speed). They watched the water molecules on the platinum surface.
• The Result: They saw that when they made the voltage more negative (like charging a battery), the water molecules flipped over. The hydrogen atoms (the "heads" of the water molecule) turned to face the metal, while the oxygen turned away.
• Why it matters: This "flipping" changes how the battery works. The AI saw this happen clearly, matching the slow, expensive super-computer results perfectly, but in a fraction of the time.

The Big Picture Takeaway

This paper is like giving scientists a universal remote control for electrochemical reactions.

• Before: You had to choose between "Fast but blurry" (old AI) or "Slow but clear" (traditional physics). You couldn't see the electricity and the movement at the same time.
• Now: This new framework gives you Fast AND Clear. It lets you watch the atomic dance and the electron flow simultaneously, under any voltage you want.

This is a huge step forward for designing better batteries, cleaner fuel cells, and more efficient chemical factories, because scientists can now simulate years of battery life or complex reactions in a matter of hours on a standard computer.

Technical Summary

1. Problem Statement

Electrochemical interfaces (solid-liquid) are critical for technologies like electrocatalysis and batteries, but their complex coupling of electrode surfaces, solvent structure, and electric potential is difficult to characterize experimentally.

• Limitations of DFT: While Density Functional Theory (DFT) is accurate, its computational cost ($O(N^3)$) prevents large-scale and long-timescale simulations required to capture rare events and statistical convergence.
• Limitations of Implicit Solvation: Standard implicit solvation models (e.g., VASPsol) often suffer from "solvent leakage," where the continuum model penetrates regions too small for actual water molecules.
• Limitations of Existing MLFFs: Current Machine Learning Force Fields (MLFFs) for electrochemical systems often require complex architectures with multiple outputs or treat electron number as a dynamical variable, leading to potential fluctuations. Crucially, existing MLFFs generally predict only atomic forces and energies, failing to predict electron density distributions under arbitrary potentials. This forces researchers to perform frame-by-frame DFT calculations for electronic analysis, creating a computational bottleneck.

2. Methodology

The authors propose a unified framework consisting of three integrated components: Data Generation, Model Training, and Model Application.

A. Data Generation: Hy DFT Software

• Tool: A new interface package called Hy DFT (Grand Canonical DFT with Hybrid Solvation Model).
• Function: It performs Constant-Potential Ab Initio Molecular Dynamics (CP-AIMD) using VASP and VASPsol++ (which uses a non-local cavity definition to fix solvent leakage).
• Algorithm: It employs a capacitance-Newton iteration strategy to iteratively update the electron number ($N_e$) until the Fermi level ($E_F$) converges to a target potential ($U$).
• Output: It automatically embeds the converged electric potential as a scalar field attribute into the atomic structure (Extended XYZ format), creating a dataset $(R, U)$ ready for ML training without post-processing.

B. Model Training: Dual-Decoupled Architecture

The framework utilizes two independent but philosophically aligned Equivariant Graph Neural Networks (GNNs) that share a core design principle: Early Fusion of Electric Potential. The electric potential is treated as a scalar input feature and concatenated with atomic species embeddings at the very first layer.

1. PE-MACE (Potential-Embedded MACE):

   • Goal: Predict atomic forces and energies (Atomic Scale).
   • Architecture: Based on the MACE (Higher-Order Equivariant Graph Neural Network) framework.
   • Mechanism: It takes atomic positions, species, and the scalar electric potential as inputs. It learns the potential energy surface (PES) under varying potentials, enabling long-timescale Constant-Potential Molecular Dynamics (CP-MLMD).
   • Output: System energy and atomic forces.

2. PE-EDP (Potential-Embedded Electron Density Prediction):

   • Goal: Predict real-space electron density distributions (Electronic Scale).
   • Architecture: An equivariant GNN that predicts expansion coefficients for Gaussian-Type Orbital (GTO) basis functions.
   • Mechanism: It uses a hybrid approach:
     • Orbital-based: Predicts coefficients for basis functions centered on atoms to capture the bulk of the density.
     • Grid-based Residual: Adds a residual correction layer operating on a sparse grid to capture fine-grained details and mitigate basis set incompleteness.
   • Output: 3D electron density $\rho(r)$ and derived properties (e.g., Bader charges).

C. Model Application

The trained models drive simulations at arbitrary target potentials without requiring further DFT calculations. PE-MACE drives the dynamics, while PE-EDP analyzes the electronic structure of the resulting trajectories.

3. Key Contributions

• Unified Framework: First framework to simultaneously predict atomic forces and electron densities under explicit constant-potential conditions, bridging atomic and electronic scales.
• Hy DFT Software: Development of a robust tool for generating high-quality CP-AIMD training data with embedded potential information, utilizing a non-local cavity model to avoid solvent leakage.
• Potential-Embedding Strategy: A novel architectural design where electric potential is fused early in the GNN as a scalar feature, allowing the model to learn potential-dependent interactions directly without complex multi-network outputs or fictitious potentiostats.
• Efficient Density Prediction: Introduction of a PE-EDP model that achieves high-fidelity electron density prediction using a sparse sampling strategy, bypassing expensive SCF iterations.

4. Results

The framework was validated using the Pt(111)/water interface as a benchmark system across three potentials ($-0.44, -0.04, +0.26$ V vs. SHE).

• PE-MACE Performance:

  • Accuracy: Achieved an Energy RMSE of 1.8 meV/atom and Force RMSE of 12.3 meV/Å on the test set.
  • Dynamics: Radial Distribution Functions (RDFs) from 10 ps CP-MLMD simulations matched CP-AIMD references perfectly.
  • Long-timescale Simulation: 4 ns CP-MLMD simulations successfully revealed potential-induced reorientation of interfacial water. As potential became more negative, water molecules shifted toward an "H-down" orientation (hydrogen pointing toward the electrode), consistent with literature but previously inaccessible via long-timescale DFT.

• PE-EDP Performance:

  • Accuracy: Achieved a Normalized Mean Absolute Error (NMAE) of 0.848% on the test set.
  • Density Comparison: Planar-averaged electron densities ($\bar{\rho}(z)$) and Bader charges showed excellent agreement with DFT references, accurately reproducing Pt layer peaks and water layer distributions.
  • Generalization: Tested on public datasets (QM9, MD, Cubic), showing competitive accuracy with state-of-the-art models like SCDP and ELECTRA, particularly for periodic systems.

5. Significance

• Computational Efficiency: By replacing frame-by-frame DFT calculations with ML inference, the framework enables nanosecond-scale simulations of electrochemical interfaces with electronic-level resolution.
• Mechanistic Insight: It provides a powerful tool to investigate reaction mechanisms, charge transfer, and solvent reorganization under operating conditions (constant potential), which are critical for designing better batteries and catalysts.
• Scalability: The decoupled nature of the models allows for flexible application; users can run dynamics with PE-MACE and analyze electronic properties with PE-EDP on the same trajectory without retraining or re-simulating.
• Open Source: The authors have released the Hy DFT software and PE-MACE source code, fostering reproducibility and further development in the electrochemical ML community.