The UMA Force Field extension lets you run UMA directly in SAMSON for interactive molecular and materials work. It is well suited for:

• Geometry refinement and exploratory relaxation.
• Qualitative force analysis during model building.
• Fast screening before higher-cost electronic-structure calculations.
• Bond-order-assisted visualization workflows (Wiberg or Mayer modes).

UMA predictions are approximations from a learned model. They are not a replacement for full high-accuracy reference methods when final numerical certainty is required.

Model Choices

• UMA-S (smaller)
  Typical use: Interactive workflows and quick iteration
  Trade-off: Faster and lighter, with lower cost per step
• UMA-M (larger)
  Typical use: More demanding scenarios where extra model capacity helps
  Trade-off: Higher memory and compute cost

Task Presets

UMA tasks configure the target DFT level of theory and domain. Choosing the right task is critical.

• OMOL - Molecules
  Best for: Organic and molecular chemistry, including charged and open-shell systems.
  Inputs: This is the only task where total charge and spin multiplicity are expected and meaningful.
  Caveats: Training data is aperiodic; use caution for periodic or strongly inorganic systems.
• OMAT - Inorganic materials
  Best for: Bulk inorganic materials workflows (e.g., crystals and materials screening).
  Inputs: Uses periodic cells.
  Caveats: Spin polarization effects are learned, but magnetic state is not user-selectable and spin-state coverage in training is limited.
• ODAC - MOFs and direct-air-capture chemistry
  Best for: CO2/H2O adsorption scenarios in MOF-like environments.
  Inputs: Uses periodic cells.
  Caveats: ODAC training is focused on CO2/H2O adsorption; extrapolation to broader chemistries (for example hydrocarbons in MOFs) may be unreliable.
• OC20 - Catalysis-oriented systems
  Best for: Surface catalysis and adsorbate-on-slab workflows.
  Inputs: Uses periodic cells.
  Caveats: No explicit solvents or oxides in core training distribution; dispersion-free RPBE-like behavior can be limiting for larger adsorbates.
• OMC - Molecular crystals
  Best for: Molecular-crystal-like organic solid-state systems.
  Inputs: No additional task-specific inputs in this interface.
  Caveats: Best suited to molecular crystals rather than general inorganic solids.

Bond Update Modes

You can choose how covalent bonds are displayed or updated during simulation:

Default mode: Covalent.

• Covalent: standard covalent perception.
• Wiberg bond order (estimated): Wiberg bond order estimated with a deep learning model.
• Mayer bond order (estimated): Mayer bond order estimated with a deep learning model.
• Off: keep existing bond graph unchanged.

Note: Wiberg and Mayer bond orders are estimated values intended for visualization and tracking. In ambiguous situations, always validate with your reference method.

Recommended Workflow

1. Select the task that matches your system class.
2. Start with UMA-S for rapid feedback; move to UMA-M if needed.
3. For OMOL, set total charge and spin multiplicity consistently.
4. For periodic tasks, verify the cell before interpretation.
5. Use UMA to narrow candidates, then confirm final results with higher-accuracy calculations.

First Run and Access

On first use, the extension prepares a Python environment and may download model assets. UMA access may require authentication or approval through Hugging Face for gated repositories.

Practical Limits and Good Practice

• Avoid extrapolating far outside chemically realistic geometries.
• Treat very high-energy or unstable configurations with extra caution.
• Check convergence behavior instead of trusting a single step.
• For publication-critical numbers, perform an independent reference calculation.

References

UMA: B. M. Wood et al., "UMA: A Family of Universal Models for Atoms" (2025).
https://arxiv.org/abs/2506.23971

Model page: https://huggingface.co/facebook/UMA

Github page: https://github.com/facebookresearch/fairchem