QML for 2D Materials Property Prediction
Built a hybrid SchNet + variational quantum circuit pipeline for bandgap prediction across 2DMatPedia. After correcting leakage and spatial-feature normalization issues, the model reached R² ≈ 0.782 and MAE = 0.439 eV on the verified run.
Can a variational quantum circuit add useful modeling capacity after a graph neural network has learned the structure of a 2D material? I explored that question through bandgap prediction on the 2DMatPedia corpus, using crystal graphs and compact physics descriptors rather than treating each material as a flat feature vector.
The project developed across three architectures to isolate quantum contributions. Model C established an amplitude-encoded Coulomb-matrix baseline. Model A introduced graph representations and angle encoding. Model B became the mainline: a periodic-boundary-aware SchNet encoder, physics features, a learned compression layer, and an RX-encoded PennyLane VQC head.
Early runs produced plausible losses while drifting toward the target mean. I tracked that behavior with a variance-ratio diagnostic and audited the full data path. The pipeline corrections included training-fold-only normalization, explicit exclusion of target-correlated features, and correct handling of spatial descriptors (std_x, std_y, std_z), establishing a verifiable training process.
The preserved corrected run reached R² ≈ 0.782 and MAE = 0.439 eV, with a variance ratio around 0.93. This is a strong single-fold result on the retained bandgap configuration, demonstrating that the hybrid model learned sample-specific variation rather than collapsing toward an average prediction.
The public repository preserves the three-model lineage, runnable source code, automated tests, representative artifacts, and the explicit reasoning behind the data-path corrections. It pairs an ambitious hybrid architecture with the rigorous validation discipline required for scientific machine learning.