Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 15·5 min read
A Fixed-Point Neural Operator for Size- and Functional-Transferable Hamiltonian Prediction
TL;DR
Predicting the Kohn-Sham Hamiltonian, essential for understanding electronic structures, has limitations with traditional methods. HamEvo, a fixed-point neural operator, learns self-consistent updates to accurately predict Hamiltonians.
✦ Why It Matters
Engineers can leverage HamEvo for faster and more accurate electronic structure predictions in large molecular systems.
Key Takeaways
- Utilize HamEvo for efficient Hamiltonian predictions in computational chemistry.
- Achieve chemical-accuracy in energy predictions with minimal data.
- Implement few-shot fine-tuning for larger molecular systems.
- Explore temperature effects on electronic properties using HamEvo.
Full Summary
In computational chemistry, predicting the Kohn-Sham Hamiltonian is crucial for accurate electronic structure calculations but often suffers from inefficiencies. HamEvo is introduced as a fixed-point neural operator that learns the single-step self-consistent field update, effectively converging to the Hamiltonian.
It is pre-trained on self-consistent trajectories and calibrated using density-matrix supervision. Benchmarks show that HamEvo reduces Hamiltonian prediction errors by 35-49% compared to traditional regression methods and predicts the highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) energies with mean absolute errors of 0.036 and 0.053 eV, respectively.
Additionally, it can be fine-tuned with only 20 reference conformations to handle larger molecules, up to 122 atoms. HamEvo also captures temperature-dependent changes in the HOMO-LUMO gap, enhancing its applicability in thermal molecular dynamics.
Overall, inference with HamEvo is up to 242 times faster than conventional density functional theory (DFT) methods.