Skip to content
Levitating Labs

Back to all threads

Experiment 02: Equivariant GNNs for Crystal Lattice Energy Prediction

Updated:

Daily experiment log evaluating E(3)-equivariant Graph Neural Networks on periodic crystal structure unit cells.

Day 1: Periodic Boundary Condition Setup

18 Sep 2026

r⃗ij=(xj+R⋅n)−xi\vec{r}_{ij} = (\mathbf{x}_j + \mathbf{R} \cdot \mathbf{n}) - \mathbf{x}_i

# Periodic edge vector calculation
def periodic_vector(pos1, pos2, lattice, cell_offsets):
    shift = torch.matmul(cell_offsets.float(), lattice)
    return (pos2 + shift) - pos1

Day 3: Convergence Monitoring

20 Sep 2026

Next Steps: Testing gradient clipping at ∥g∥≤1.0\|g\| \le 1.0 and adding radial bessel basis functions to smooth atomic pair interactions at cutoff radius rc=6.0 A˚r_c = 6.0\text{ \AA}.


Day 4: Current Status 🧪

21 Sep 2026

Previous ThreadExperiment 01: Out-of-Domain Generalization of ESM2 Embeddings in Protein-Ligand Binding