OrbNet: Deep learning for quantum chemistry using symmetry-adapted atomic-orbital features.

Area of Science:

• Computational chemistry
• Machine learning in quantum mechanics
• Materials science

Background:

• Accurate prediction of molecular energies is crucial for drug discovery and materials design.
• Existing methods often require significant computational resources.
• Developing efficient and accurate predictive models is an ongoing challenge.

Purpose of the Study:

• To introduce OrbNet, a novel machine learning method for predicting energy solutions from the Schrödinger equation.
• To demonstrate OrbNet's superior learning efficiency and transferability compared to existing methods.
• To achieve density functional theory (DFT) accuracy at a significantly reduced computational cost.

Main Methods:

• Utilized symmetry-adapted atomic orbital features.
• Employed a graph neural-network architecture (OrbNet).
• Leveraged low-cost features from semi-empirical electronic structure calculations.

Main Results:

• OrbNet outperforms existing methods in learning efficiency and transferability for DFT predictions.
• Achieved chemical accuracy comparable to DFT.
• Demonstrated applicability to diverse molecular datasets, including drug-like molecules and conformer benchmarks.

Conclusions:

• OrbNet offers a highly efficient and accurate approach for predicting molecular energies.
• The method significantly reduces computational cost (by 1000-fold or more).
• OrbNet holds promise for accelerating drug discovery and materials science research.