psiformer_torch

PsiFormer Checkpoint: Hydrogen → Oxygen

This repository contains pretrained PsiFormer checkpoints for electronic-structure modeling across atomic systems ranging from Hydrogen (Z=1) to Oxygen (Z=8).

The model is designed for variational quantum Monte Carlo (VMC)–style wavefunction modeling, with a Transformer-based architecture that captures electron–electron correlations efficiently and scalably.

---

Model Overview

• Architecture: PsiFormer (Transformer-based wavefunction ansatz)
• Task: Electronic wavefunction approximation
• Method: Variational Monte Carlo (VMC)
• Atomic range: Hydrogen → Oxygen
• Framework: PyTorch
• Precision: FP32 (unless otherwise specified)

The model outputs parameters of a many-body wavefunction that can be used to estimate ground-state energies and other observables via Monte Carlo sampling.

---

Training Details

• Systems: Isolated atoms with atomic numbers Z = 1–8
• Electrons: Corresponding neutral configurations
• Optimization: Stochastic gradient–based optimization of variational energy
• Sampling: Metropolis–Hastings MCMC
• Objective: Minimize the expectation value of the Hamiltonian

Exact hyperparameters (learning rate, batch size, number of walkers, etc.) should be considered checkpoint-specific and are documented in the accompanying configuration files when available.

---

Intended Use

This checkpoint is intended for:

• Initializing PsiFormer models for light atoms
• Transfer learning to larger atoms or small molecules
• Benchmarking neural quantum states
• Research and educational purposes in computational quantum physics

It is not intended for production chemistry workflows without further validation.

---

Example Usage

import torch
from psiformer import PsiFormer

model = PsiFormer(...)
state_dict = torch.load("psiformer_h_to_o.pt", map_location="cpu")
model.load_state_dict(state_dict)
model.eval()

Refer to the PsiFormer repository for full examples including sampling and energy evaluation.

---

Limitations

• Trained only on isolated atoms, not molecules
• Accuracy degrades outside the Z = 1–8 range
• Performance depends strongly on sampling quality and optimization setup
• No relativistic or spin–orbit effects included

---

Citation

If you use this checkpoint in academic work, please cite the corresponding PsiFormer paper or repository.

@misc{psiformer,
  title={PsiFormer: Transformer-based Neural Quantum States},
  author={...},
  year={202X}
}

---

License

Specify the license here (e.g. MIT, Apache 2.0, custom research license).

---

Contact

For questions, issues, or collaborations, please open an issue in the main PsiFormer repository.