DefectNet: MIT AI Model Trained on 2,000 Semiconductors Detects Six Coexisting Substitutional Defects

Dataset Construction

The researchers built a dataset of 16,000 doped super‑cell spectra derived from 2,000 perfect crystal materials, covering binary to five‑component semiconductors. Substitutional defects were sampled from the first 56 elements of the periodic table, either singly or in combination, to emulate realistic co‑doping scenarios. A machine‑learning‑based dopant recommender selected n‑type and p‑type candidates, after which each super‑cell (433–500 atoms) was relaxed until forces fell below 0.01 eV/Å. Finite‑displacement calculations yielded phonon density of states (PDoS) curves, which were Gaussian‑smoothed to mimic experimental resolution.

DefectNet Architecture

DefectNet is implemented in PyTorch and consists of four modular components:

Spectral encoder : a 1‑D convolutional network processes three length‑100 inputs (host PDoS, doped PDoS, and host composition vector) as a three‑channel signal, producing 100 spectral tokens of 128 dimensions each.

Dopant embedding : a 56‑dimensional binary vector indicating possible dopants is projected to a 128‑dimensional embedding that serves as the global query for attention.

Multi‑head attention : the dopant embedding queries the spectral token values, allowing the model to focus on defect‑relevant vibrational features.

Dopant‑masked output : predictions for the 56 elements are hard‑masked so that only candidate dopants receive non‑zero outputs, improving training stability and respecting physical priors.

The model accepts four inputs: host composition, host PDoS, doped PDoS, and an initial dopant guess (either heuristic or generated by the recommender). Although trained on simulated data, the framework can be fine‑tuned on experimental spectra such as inelastic neutron scattering.

Performance on Simulated Spectra

DefectNet was evaluated on binary (SiC, AlAs) and ternary (AgGaS₂, InCuSe₂) semiconductors. Even at low dopant levels (~1 %), the model captured subtle vibrational changes and recovered accurate concentrations. For more complex ternary systems, it reliably tracked PDoS variations and inferred dopant levels, demonstrating robustness to multiple inequivalent atomic sites.

In a comprehensive test with up to six co‑existing substitutional defects, concentrations as low as 0.2 % were correctly identified. The model maintained high fidelity on in‑distribution crystals (those seen during training) and showed only modest accuracy loss on out‑of‑distribution crystals, while effectively suppressing predictions for spurious “interfering” defects.

Experimental Fine‑tuning

To assess real‑world applicability, the authors fine‑tuned DefectNet on experimental data from SiGe alloys. A training set of 100 amorphous Si super‑cells (generated via Si‑GAP‑18 quench simulations) spanned disorder levels and Ge concentrations from 0 % to 25 %. After fine‑tuning, the model achieved a root‑mean‑square error of 0.019 on the test set and predicted Ge concentrations of 7 %, 13 %, and 22 % for experimental samples with nominal x = 5 %, 10 %, and 20 %, respectively. Similar success was reported for Al‑doped MgB₂ superconductors up to 25 % dopant levels.

Challenges and Outlook

Key challenges include diminishing vibrational signatures at ultra‑low defect concentrations, which can be masked by noise, and the current limitation to substitutional defects only. Extending the approach to interstitials, vacancies, Frenkel pairs, or defect clusters would broaden its utility. While simulated data provide strong generalization, further experimental fine‑tuning remains essential, and a truly plug‑and‑play model for raw experimental spectra is a future goal.

Overall, DefectNet represents a significant step toward a unified, data‑driven paradigm for non‑destructive defect identification, offering a pathway for automated, interpretable defect engineering in complex materials.