Multi-fidelity Machine Learning Interatomic Potentials for Charged Point Defects

This paper introduces a multi-fidelity machine learning framework incorporating global defect charge embeddings to accurately model charged point defects in Sb2Se3, enabling the prediction of stable structures and charge-transition levels at a fraction of the computational cost of direct quantum mechanical calculations.

The Gist

Imagine you are trying to find the perfect spot to park a car in a massive, complex city. The city represents a material (like a semiconductor), and the "parking spot" is the most stable, energy-efficient arrangement of atoms.

For decades, scientists have used a very precise but incredibly slow method (like a super-accurate GPS that takes hours to calculate a route) to find these spots. Recently, they started using Machine Learning Interatomic Potentials (MLIPs)—think of these as "AI driving assistants" trained on millions of normal, everyday driving scenarios (perfect crystals) to predict the best route instantly.

However, this paper reveals a major problem: These AI assistants fail miserably when the car has a flat tire or is carrying a heavy, unusual load. In scientific terms, they can't handle "defects" (missing atoms) or "charged states" (extra or missing electrons) inside the material.

Here is a breakdown of what the authors did to fix this, using simple analogies:

1. The Problem: The AI is "Blind" to Charge

The current generation of AI driving assistants was trained only on "perfect" cars driving on "perfect" roads.

• The Flaw: When you introduce a defect (a missing atom) or change the charge (add electrons), the physics of the material changes completely. It's like the AI thinks a car with a flat tire should drive exactly the same as a car with full tires.
• The Result: The AI predicts the wrong parking spot. It might tell you to park in a spot that looks okay but is actually unstable, or it misses the true best spot entirely. In the paper, they tested this on a material called $Sb_2Se_3$ (used in solar cells) and found the AI was consistently wrong about where the atoms should settle.

2. The First Fix: Giving the AI "Charge Glasses"

The authors realized the AI didn't know which version of the car it was driving. Is it the neutral version? The positively charged version? The negatively charged version?

• The Solution: They added "Global Charge Embeddings."
• The Analogy: Imagine giving the AI driver a pair of special glasses that change color depending on the car's load. If the car is "charged," the glasses turn red; if it's neutral, they turn blue. Now, the AI knows, "Ah, this is a heavy-load scenario; I need to drive differently."
• The Outcome: With these "glasses," the AI could finally distinguish between different charged states. It stopped guessing and started predicting the correct atomic structure with near-perfect accuracy (within 0.05 Angstroms, which is smaller than an atom).

3. The Second Fix: The "Multi-Fidelity" Strategy

Even with the new glasses, training the AI to be perfect is expensive. To get the highest accuracy, you usually need to run super-slow, high-precision calculations (like hiring a team of expert engineers to check every inch of the car).

• The Problem: You can't afford to hire expert engineers for every single test drive.
• The Solution: They used a Multi-Fidelity (MF) approach.
• The Analogy: Imagine you want to find the best route through a city.
  • Low-Fidelity: You use a cheap, fast map app (like Google Maps on a basic phone) to scan the whole city quickly. It's fast but might miss a tiny, perfect shortcut.
  • High-Fidelity: You hire a local expert to check the top 10 most promising shortcuts.
  • The Magic: The AI learns the difference between the "cheap map" and the "expert's advice." It learns to take the cheap map's speed but apply the expert's corrections.
• The Outcome: This allowed them to find the true best parking spot (the global minimum) that standard methods missed, but they did it 1,000 times faster than before.

Why Does This Matter?

Defects in materials are like the "secret sauce" that makes solar cells work, batteries last longer, or catalysts clean the air.

• Before: Scientists had to guess where these defects were or spend years simulating them.
• Now: With this new "Charge-Aware, Multi-Fidelity AI," scientists can instantly find the most stable structures and predict how these materials will behave, all while using a fraction of the computer power.

In a nutshell: The authors built a smarter AI that knows how to handle "broken" or "charged" materials and taught it to learn from cheap, fast data while only paying for expensive, high-quality checks when absolutely necessary. This opens the door to designing better solar panels, batteries, and electronics much faster than ever before.

Technical Summary

Here is a detailed technical summary of the paper "Multi-fidelity Machine Learning Interatomic Potentials for Charged Point Defects."

1. Problem Statement

Machine Learning Interatomic Potentials (MLIPs) have achieved high accuracy in reproducing the properties of bulk materials. However, their application to charged point defects in semiconductors and insulators remains a significant bottleneck due to two primary challenges:

• Data Scarcity and Domain Gap: Current "foundation" MLIPs are trained predominantly on pristine bulk structures. They fail to capture the complex, symmetry-breaking local environments and electronic reconstructions specific to defects.
• Charge-State Indistinguishability: Standard MLIPs rely on locality assumptions and descriptors that capture only atomic species and local geometry. They lack explicit information about the global charge state. Consequently, different charge states of the same defect (which possess distinct electronic occupations and local relaxations) are mapped onto a single energy landscape. This leads to the collapse of charge-dependent Potential Energy Surfaces (PESs), causing models to predict incorrect ground-state structures or miss metastable configurations entirely.
• Computational Cost: Accurate defect modeling requires high-level electronic structure theory (e.g., hybrid functionals like HSE06) to correct bandgap errors and delocalization issues inherent in semi-local functionals (e.g., PBE). However, the computational cost of hybrid functional calculations makes exhaustive structure searching (e.g., finding global minima) prohibitively expensive.

2. Methodology

The authors propose a two-pronged approach to overcome these limitations using the antimony vacancy ($V_{Sb}$) and selenium vacancy ($V_{Se}$) in $Sb_2Se_3$ as case studies.

A. Global Defect Charge Embeddings

To address the charge-state indistinguishability, the authors developed a bespoke MLIP architecture based on MACE (a many-body equivariant message-passing model).

• Architecture Modification: They introduced a global charge embedding mechanism. The total defect charge ($q_{tot}$) is mapped to a learnable vector and added to the atomic species embeddings during feature initialization.
• Dual Pathway Learning:
  1. Interaction Pathway: The charge information propagates through message-passing layers, allowing the network to learn charge-dependent interaction energies ($E_{inter}$).
  2. Embedding Pathway: The charge-enriched embedding is read out directly to provide a geometry-independent energy contribution ($E_{emb}$).
• Training Data: The model is trained on DFT data covering multiple charge states ($q = -3$ to $+1$) and diverse bond-distorted configurations generated via the ShakeNBreak algorithm.

B. Multi-Fidelity (MF) Training Strategy

To reduce the computational cost of generating high-quality training data while maintaining accuracy, the authors implemented a multi-fidelity approach:

• Low-Fidelity Data: A large volume of data is generated using low-cost PBE (semi-local functional) calculations with converged k-point sampling to explore the PES broadly.
• High-Fidelity Corrections: A sparse subset ($\sim$10%) of the PBE-relaxed structures is selected via stratified sampling and re-calculated using high-quality HSE06 (hybrid functional) single-point energies.
• $\Delta$-Learning: The model learns the functional-dependent difference ($\Delta$) between HSE06 and PBE. It utilizes a shared latent representation with separate readout channels for low-fidelity (PBE) and high-fidelity (HSE06) outputs. This allows the model to correct the PBE PES to approximate HSE06 accuracy at a fraction of the cost.

3. Key Results

Failure of Foundation Models

• Benchmarking four state-of-the-art foundation models (MACE-MH-1, MACE-MPA-0, GRACE-2L-OMAT, MatterSim) against PBE-DFT revealed they failed to identify correct defect ground states.
• Structural Root Mean Square Deviations (RMSD) ranged from 0.2 to 0.4 Å, far exceeding the 0.1 Å threshold for reliability.
• The models tended to relax defects into high-symmetry, bulk-like configurations or spurious local minima, missing critical electronic reconstructions (e.g., the formation of Se trimers in neutral $V_{Sb}$).

Success of Charge-Embedded Models

• Structural Accuracy: The bespoke charge-embedded model successfully identified ground-state configurations for all five charge states with a global RMSD < 0.05 Å.
• Energetic Accuracy: The model predicted thermodynamic charge transition levels with a maximum deviation of ~0.01 eV compared to HSE06 reference calculations.
• Descriptor Separation: Principal Component Analysis (PCA) showed that without charge embeddings, different charge states overlapped in descriptor space. With embeddings, the model learned distinct, charge-specific clusters while retaining sensitivity to local structural variations.
• Metastable Detection: The model could distinguish metastable configurations with energy differences as small as 1 meV/atom.

Efficiency of Multi-Fidelity Approach

• Discovery of Global Minima: Standard workflows (even with coarse HSE06 sampling) often identified only metastable minima. The MF-trained model discovered a distinct global minimum 0.02 eV lower in energy, which was subsequently confirmed by full HSE06 relaxation.
• Computational Savings: The MF strategy reduced the computational cost of comprehensive PES exploration by approximately three orders of magnitude compared to direct hybrid functional searches.
• Surrogate Capability: The trained MF model acted as an efficient surrogate, enabling the rapid calculation of Configuration Coordinate Diagrams (CCDs) and other properties (e.g., carrier capture rates) at hybrid-level accuracy.

4. Significance and Impact

• Bridging the Accuracy-Cost Gap: The work demonstrates a practical framework to achieve ab initio (hybrid functional) accuracy for charged defects at a computational cost comparable to semi-local DFT, making high-throughput defect screening feasible.
• Overcoming the "Out-of-Distribution" Limit: It proves that foundation models trained solely on bulk data are insufficient for defect physics and that global charge conditioning is essential for generalizing MLIPs to charged systems.
• Foundation for Future Models: This study lays the groundwork for developing universal "defect foundation models" capable of generalizing across different materials, charge states, and levels of theory, accelerating the discovery of functional materials for photovoltaics, solid-state electrolytes, and catalysis.

Conclusion

The paper establishes that combining global charge embeddings with a multi-fidelity training strategy solves the critical challenges of modeling charged point defects. This approach enables the reliable identification of global energy minima and accurate prediction of charge transition levels, overcoming the limitations of both current foundation models and traditional high-cost DFT workflows.