2. What is a Migration Barrier?

Lithium diffusion occurs via hopping between two equilibrium lattice sites.

Consider:

• An initial site with energy Einitial 

• A final site with similar equilibrium energy

• A saddle point along the minimum energy path
  
  

The migration barrier is defined as:

Ea =Esaddle − Einitial 

Graphically, this corresponds to the energy difference between the lowest energy configuration (stable site) and the highest energy configuration along the minimum energy path. 

Physically, the saddle point represents the most constrained configuration during hopping, where:

• Li ion passes through a bottleneck

• Electrostatic repulsion is maximal

• Local coordination is distorted
  
  

The magnitude of this barrier determines how easily lithium can move through the material.

3. How is the Barrier Computed?

The Nudged Elastic Band (NEB) Method

To compute Ea​, one must determine the minimum energy path (MEP) between initial and final configurations. The standard method is the Nudged Elastic Band (NEB) method:

1. Generate a sequence of intermediate images between initial and final states.

2. Relax these images while constraining them along the path.  The force perpendicular to the path is 0.

3. Identify the highest-energy image — the saddle point.
   
   

When forces and energies are computed with Density Functional Theory (DFT), this becomes DFT-NEB.

However, DFT-NEB is computationally expensive because:

• Each image requires a full DFT calculation.

• A single hop may require 5–9 images.

• Supercells often contain 100+ atoms.
  
  

Screening thousands of materials becomes prohibitive.

4. Approximate Physics: BVSE

To overcome DFT cost, approximate methods like Bond Valence Site Energy (BVSE) can estimate migration barriers cheaply. 

BVSE:

• Uses empirical electrostatic models.

• Estimates site energies from bond valence parameters.

• Provides rapid barrier estimates.

Advantages:

• Fast

• Scalable to 100k+ structures
  
  

Limitations:

• Approximate physics

• Less accurate than DFT

• Systematic biases
  
  

Thus, BVSE provides large-scale but imperfect data.

5. The Central Scientific Question

Can large-scale approximate physics be used to learn transferable representations that improve high-fidelity predictions?

Formally:

Given:

• Large dataset DBVSE (approximate physics)

• Small dataset DDFT (ground truth)
  
  

We ask:

Does learning on DBVSE improve generalization on DDFT?

6. Why Graph Neural Networks?

Migration barriers are not simple functions of composition alone. They depend on:

• Local atomic coordination

• Geometry of diffusion bottleneck

• Bond lengths and distortions

• Electrostatic environment
  
  

These are inherently structural and geometric properties.

We therefore represent each migration hop as a graph:

• Nodes: Atoms in the supercell

• Special centroid atom (“X”) marks bottleneck location

• Edges: Neighbor interactions within cutoff

• Node features: Atomic number + distance to centroid

• Edge features: Interatomic distances
  
  

This allows the model to learn local geometric environments directly.

7. Transfer Learning Hypothesis

We hypothesize:

1. Pretraining on 122k BVSE barriers allows the model to learn structural migration physics.

2. Fine-tuning on 1.6k DFT barriers adjusts energy scale and corrects systematic bias.

3. The resulting model generalizes better than training on DFT alone.
   
   

This approach leverages:

• Abundant approximate physics

• Scarce high-fidelity quantum data

8. Why This Matters

If successful, this framework enables:

• Rapid screening of thousands of battery materials.

• Reduced reliance on expensive DFT-NEB.

• Scalable discovery of solid electrolytes.

• Data-efficient learning in materials science.
  
  

More broadly, this project demonstrates large-scale approximate simulations can serve as effective pretraining corpora for high-fidelity quantum predictions.