NEPMaker: Active learning of neuroevolution machine learning potential for large cells

The paper introduces NEPMaker, a D-optimality-driven active learning framework integrated with the GPUMD package that enables the efficient construction of robust and transferable neuroevolution potentials for large-scale simulations of complex materials by embedding extrapolative atomic environments into locally periodic structures to minimize labeling costs.

The Gist

Imagine you are trying to teach a robot chef how to cook a perfect meal. You start by showing it a few recipes (the training data). The robot learns to cook those dishes perfectly. But what happens if you ask it to cook a dish it has never seen before, or if the ingredients are slightly different? The robot might guess wildly, burn the food, or even break the stove. In the world of computer simulations, this "robot" is a Machine Learning Potential (MLP), and the "food" is a simulation of atoms moving around.

The problem is that scientists want to simulate huge, complex systems (like a melting piece of metal or a crystal changing shape), but they can't afford to calculate every single atom's behavior using the most accurate (but incredibly slow) physics methods. They need the robot chef to be fast and accurate.

Here is how the paper NEPMaker solves this problem, explained simply:

1. The Problem: The "Blind Spot"

Standard AI models are great at what they've seen, but terrible at what they haven't. If a simulation encounters a weird atomic arrangement the AI hasn't memorized, it might crash or give nonsense results.

• The Old Way: To fix this, scientists would stop the simulation, find the weird part, run a super-slow, perfect physics calculation on the entire huge system to get the right answer, and then restart. This is like stopping a movie to re-film the whole scene just because one actor forgot a line. It's too expensive and slow.

2. The Solution: The "Spotlight" (Active Learning)

The authors created a tool called NEPMaker. Think of it as a smart spotlight that scans the simulation while it's running.

• The D-Optimality Criterion: This is the "spotlight's brain." It constantly asks: "Hey, is this atom doing something weird that we haven't seen before?"
• If the answer is No, the simulation keeps running fast.
• If the answer is Yes (the robot is in its "blind spot"), the system pauses only that specific part to get the correct answer from the slow, perfect physics engine.

3. The Innovation: The "Micro-Island" Strategy

Here is the tricky part. If you find a weird atom in a giant simulation of a million atoms, you can't just cut that atom out and put it in a vacuum (empty space) to study it. That's like taking a fish out of the ocean and putting it in a bowl of air; it dies, and the data is useless. The surrounding atoms matter!

NEPMaker's clever trick:
Instead of cutting out a weird atom and leaving it alone, they cut out a small "island" of atoms around it. Then, they use a special optimization technique to gently adjust the "shoreline" (the boundary atoms) of this island.

• The Analogy: Imagine you are studying a specific tree in a dense forest. Instead of cutting the tree down and putting it in a desert (which changes how it looks), you cut out a small patch of forest. You then gently trim the bushes at the edge of the patch so they look like a normal forest edge, ensuring the tree inside is still in a realistic environment.
• This ensures the "weird" atom is studied in a realistic setting, making the data accurate without needing to simulate the whole million-atom system.

4. The Result: A Self-Improving Loop

The process works like a video game with an auto-upgrade system:

1. Run: The simulation runs.
2. Spot: NEPMaker spots a "weird" atomic environment.
3. Fix: It extracts that small "island," fixes the edges, and runs a perfect calculation on just that island.
4. Learn: It adds this new, perfect data to the robot's recipe book.
5. Repeat: The simulation continues, now smarter, until it never gets lost again.

Real-World Examples in the Paper

The authors tested this on three different "kitchens":

1. Melting Sodium: They simulated a block of sodium metal melting. The AI learned to handle the transition from solid to liquid perfectly, predicting the melting point almost exactly like real life.
2. CsPbI₃ (A Solar Cell Material): This material changes shape as it gets hot. The AI successfully tracked these shape-shifting transitions, which are very hard to predict.
3. Gallium Nitride (GaN): They simulated a massive crystal changing its internal structure under pressure. Because the system was so huge, the AI had to learn new "moves" on the fly, which it did successfully using this method.

Why This Matters

NEPMaker is a game-changer because it allows scientists to simulate huge, complex materials with near-perfect accuracy without waiting years for the computer to finish. It turns the simulation into a self-teaching system that gets better the more it runs, bridging the gap between "fast but dumb" and "slow but smart."

In short: It's a smart robot that knows when it's confused, stops just long enough to ask for help on the specific confusing part, and then keeps going, getting smarter every time.

Technical Summary

1. Problem Statement

Machine Learning Potentials (MLPs), particularly Neural Network-based potentials like Neuroevolution Potential (NEP), offer near first-principles accuracy with computational efficiency suitable for large-scale Molecular Dynamics (MD) simulations. However, a critical limitation remains: extrapolation failure. When MD simulations explore atomic environments outside the training distribution (e.g., defects, phase transitions, or high-temperature states), MLPs often produce unreliable predictions, leading to simulation instability or incorrect physical behavior.

Existing solutions face significant hurdles:

• Uncertainty Quantification (UQ): Traditional methods like ensemble learning are computationally expensive (training multiple models), while Bayesian methods (e.g., Gaussian Process Regression) suffer from cubic scaling with dataset size, making them unsuitable for large datasets.
• Active Learning in Large Cells: Standard active learning requires labeling entire simulation configurations with expensive Density Functional Theory (DFT) calculations. For large supercells (thousands to millions of atoms), this is prohibitively costly.
• Fragment Extraction Issues: Extracting local atomic environments from large cells often involves placing them in vacuum (non-periodic), which creates unphysical boundary effects and ambiguous energy assignments. Conversely, relaxing surrounding atoms via DFT is accurate but computationally heavy.

2. Methodology

The authors propose NEPMaker, a D-optimality-driven active learning framework integrated into the GPUMD package. The methodology consists of four core components:

A. D-Optimality for Uncertainty Quantification

Instead of training multiple models, NEPMaker uses the D-optimality criterion to estimate uncertainty on-the-fly.

• Mechanism: It constructs a design matrix based on the descriptor vectors (radial and angular parts) of the training set.
• Extrapolation Grade ($\gamma$): By selecting an "active set" of maximally linearly independent descriptors using the MaxVol algorithm, the system calculates an extrapolation grade for any new atomic environment.
  • $\gamma \leq 1$: The environment is within the interpolation range (reliable).
  • $\gamma > 1$: The environment is an extrapolation (unreliable/high uncertainty).
• Efficiency: This requires only a single model evaluation, making it highly efficient for GPU-accelerated MD.

B. Periodic Fragment Extraction with Boundary Optimization

To enable active learning in large cells without labeling the whole system, NEPMaker extracts local atomic environments but avoids the "vacuum cluster" problem.

• Strategy: Instead of placing the extracted fragment in vacuum, it constructs a periodic cell.
• Boundary Optimization: The atoms outside the target local environment (the boundary) are optimized to minimize an uncertainty-based objective function.
  • Unlike previous methods (e.g., IDEAL) that fix the lattice or use different descriptors, NEPMaker optimizes both lattice parameters and atomic positions simultaneously.
  • It uses a hard-sphere potential to prevent boundary atoms from intruding into the target region.
  • Goal: Ensure the surrounding environment remains physically meaningful and close to the training distribution, preventing artificial extrapolation at the boundaries while preserving the target environment.

C. Multi-Element Handling

For multi-component systems, the magnitude of descriptor vectors ($B$ vectors) can vary significantly between elements. NEPMaker partitions atoms by chemical species and applies the MaxVol algorithm independently to each group to prevent selection bias toward specific elements.

D. Automated Workflow

The framework operates in an iterative loop:

1. Initial Training: Fit a baseline NEP.
2. Exploration: Run MD simulations; identify configurations with high $\gamma$.
3. Selection & Refinement: Extract local environments, optimize boundaries, and select informative structures.
4. Labeling: Perform DFT calculations on the refined fragments.
5. Retraining: Update the training set and retrain the NEP.
6. Convergence: Repeat until no new extrapolative environments are detected.

3. Key Contributions

1. NEPMaker Framework: The first implementation of D-optimality-based active learning specifically for NEP within the GPUMD package, enabling scalable, on-the-fly potential refinement.
2. Scalable Large-Cell Strategy: A novel method to incorporate large-scale MD simulations into active learning by extracting and optimizing local periodic fragments, bypassing the need to label entire supercells with DFT.
3. Robust Boundary Handling: Introduction of a simultaneous lattice and atomic position optimization for extracted fragments, ensuring physical consistency and reducing extrapolation errors at boundaries without expensive DFT relaxation of the entire fragment.
4. Efficient Uncertainty Metric: Demonstration that the D-optimality extrapolation grade ($\gamma$) is a reliable, computationally cheap indicator of model uncertainty, outperforming ensemble methods in efficiency and Bayesian methods in scalability.

4. Results and Validation

The framework was validated on three distinct systems:

• Sodium (Na) Melting:
  • Task: Simulate melting from solid to liquid and back.
  • Outcome: The active learning loop converged in 7 iterations (130 structures). The resulting NEP accurately predicted the melting point (~350 K, close to the experimental 370 K) using a 20,000-atom two-phase simulation.
• CsPbI$_3$ Solid-Solid Phase Transitions:
  • Task: Capture temperature-driven transitions between orthorhombic ($\gamma$), tetragonal ($\beta$), and cubic ($\alpha$) phases.
  • Outcome: Converged in 23 iterations (401 structures). The potential successfully reproduced the correct sequence of phase transitions and lattice parameter evolution in a 23,040-atom supercell, matching known transition temperatures.
• GaN B4-B1 Phase Transition:
  • Task: Investigate size-dependent phase transitions using Metadynamics.
  • Outcome: Demonstrated the ability to handle large supercells (27,648 atoms). The method captured complex transition pathways (hexagonal ring reconstruction vs. c-axis compression) and nucleation mechanisms that are inaccessible to standard first-principles MD. The process converged in 9 iterations for the large cell.

Performance Metrics:

• The final NEP models achieved low Root Mean Square Errors (RMSE) for energy, forces, and virials across all systems.
• The method successfully reduced the number of required DFT calculations by focusing only on high-uncertainty local environments.

5. Significance

NEPMaker addresses a major bottleneck in computational materials science: the trade-off between simulation scale and model reliability.

• Automation: It enables the automated construction of reliable MLPs from scratch, even for complex processes involving defects, interfaces, and phase transitions.
• Efficiency: By decoupling the extraction of local environments from the cost of labeling entire large cells, it makes high-fidelity simulations of macroscopic systems feasible.
• Generalizability: The framework provides a scalable route for constructing transferable potentials, ensuring that MD simulations remain stable and physically accurate even when exploring uncharted configuration spaces.

This work effectively bridges the gap between first-principles accuracy and large-scale simulation capabilities, offering a robust tool for studying complex dynamical processes in materials.