Multi-Modal Foundation · sciai-lab
BOA — Basis Overlap Architecture
An equivariant graph neural network predicting molecular ground-state electron densities using basis overlap matrices.
What BOA — Basis Overlap Architecture is
BOA (Basis Overlap Architecture) is an equivariant graph neural network developed by sciai-lab to predict ground-state electron densities for molecular systems. Electron density is a fundamental quantity in quantum chemistry that governs molecular properties and interactions.
Unlike standard approaches that treat basis coefficients as generic node features, BOA incorporates the overlap matrix of the basis functions directly into its message-passing mechanism. Checkpoints are provided for common computational benchmarks, including configurations evaluated on QM9 and various molecular dynamics datasets.
How it works
BOA represents molecules as graphs and applies an equivariant architecture to predict density coefficients. Its message passing explicitly leverages the overlap matrix of the underlying basis functions that describe the electronic density, embedding functional geometric information directly into network operations.
The model is trained using an exponential moving average (EMA) of its weights. Available configurations include large and small variants evaluated on PySCF and VASP generated datasets, alongside specialized models with reduced interaction cutoffs for molecular size extrapolation.
Key facts
- Architecture
- Equivariant graph neural network (BOA)
- Framework
- PyTorch
- Evaluation datasets
- QM9 (PySCF, VASP) and MD trajectories (benzene, resorcinol, phenol, malonaldehyde, ethanol, ethane)
What you can use it for
- Ground-state electron density prediction
- Predicting electronic charge densities of small organic molecules from datasets such as QM9.
- Molecular dynamics density tracking
- Estimating ground-state electron densities across conformations in molecular dynamics trajectories for molecules including benzene, phenol, and ethanol.
- Out-of-distribution size extrapolation
- Applying reduced-cutoff checkpoints to extrapolate electronic density predictions to molecules considerably larger than those encountered during training.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Achieves low normal mean absolute error (NMAE) down to 0.106% on QM9 PySCF benchmarks.
- Explicitly incorporates basis function overlap matrices into message passing rather than relying on generic node features.
- Provides a specialised reduced-cutoff variant designed specifically for extrapolation to larger molecular structures.
Watch out for
- Checkpoints are specialized to specific molecular dynamics trajectories or the chemical space represented by QM9.
- Extrapolation to substantially larger systems requires using models trained with tailored interaction cutoffs.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · lgpl-3.0
The lgpl-3.0 licence carries terms worth reading before you depend on it.
Questions
Can I use BOA — Basis Overlap Architecture commercially?
Check the licence before commercial use. The lgpl-3.0 licence carries terms worth reading before you depend on it.
What basis representation does BOA use?
BOA represents electron density through basis functions and explicitly uses their overlap matrix during message passing.
How are the weights stored in the public checkpoints?
The published checkpoints directly store the exponential moving average (EMA) weights within their state dicts, meaning no additional EMA copy step is required at inference time.
Which molecular systems have dedicated molecular dynamics checkpoints?
Checkpoints are provided for benzene, resorcinol, phenol, malonaldehyde, ethanol, and ethane.
How to cite it
A Function-Centric Graph Neural Network Approach For Predicting Electron Densities, ICLR 2026, https://openreview.net/forum?id=HDdkFjFEZd
Written from BOA — Basis Overlap Architecture's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.