Quantum software company Terra Quantum and automotive and industrial supplier Schaeffler Technologies AG & Co. KG have published peer-reviewed research in Computational Materials Science detailing the integration of hybrid quantum-classical machine learning into interatomic potential models for industrial materials simulation.
The study evaluates architectural modifications to the Allegro machine learning interatomic potential (MLIP) model, an E(3)-equivariant graph neural network designed to predict atomic energies and forces at accuracy levels approaching Density Functional Theory (DFT) while significantly reducing computational overhead. The research team constructed and benchmarked two modified Allegro variants: a classical multi-layer perceptron (MLP) expansion and a quantum-classical hybrid model incorporating Quantum Depth-Infused (QDI) variational quantum circuit layers.
| [ Benchmark Comparison Across MLIP Model Variants ] | |||
|---|---|---|---|
| Dataset & Target System | Allegro Baseline (Force MAE) | Allegro+MLP Classical (Force MAE) | Allegro+QDI Hybrid (Force MAE) |
| • Copper-Lithium (Cu-Li, 11,635 structures) | • 45.9 meV/Å | • 43.3 meV/Å | • 37.8 meV/Å (~13% error reduction) |
| • rMD17-Aspirin (Organic Molecule) | • 10.2 meV/Å | • 9.0 meV/Å | • 9.9 meV/Å (Transferred Hyperparameters) |
| • rMD17-Benzene (Organic Molecule) | • 0.43 meV/Å | • 0.28 meV/Å | • 0.30 meV/Å (Transferred Hyperparameters) |
| • Optimization Objective | • Single-objective baseline | • Multi-objective Pareto (MAE vs. Inference) | • QDI Layer Data Re-uploading Strategy |
Hyperparameter tuning across accuracy and inference time objectives was executed using a multi-objective optimization implementation (SAMO-COBRA). On the self-generated, proprietary copper-lithium dataset—comprising 11,635 DFT-calculated interfacial, slab, and melt-quench structures relevant to battery anode research—the hybrid quantum-classical variant achieved the highest force prediction accuracy, outperforming its classical MLP counterpart by approximately 13%.
Furthermore, when tested on public organic benchmarks (rMD17 aspirin and benzene), the hybrid model retained competitive accuracy even when using QDI hyperparameters transferred directly from the Cu-Li system without dataset-specific re-tuning.
Review the full open-access journal article in Computational Materials Science here and read the announcement on the Terra Quantum Portal here.
October 15, 2026