uma-ase 0.9.8

UMA MLIPs integrated with ASE workflows

uma-ase

uma-ase combines UMA machine-learned interatomic potentials with ASE workflows for:

• single-point energies
• geometry optimization
• vibrational / thermochemical analysis
• transition-state search with ASE's dimer method
• molecular dynamics

It can be used from the command line or through the bundled webapp.

Overview

• One CLI for sp, geoopt, freqs, ts, and md
• Webapp for structure loading, editing, job submission, visualization, analysis, and ioChem-BD interaction
• Persistent FIFO job queue stored server-side in jobs.jsonl
• Reproducible .use recipe files for every submitted job
• Automatic CML export after successful runs
• Live log streaming and MD trajectory capture

Citation

If results obtained with uma-ase contribute to a publication, please cite the UMA-ASE software paper:

• C. Bo, G. Dias Da Silva, F. Hosseini, A. R. Mullukkandy, H. Salazar-Lozas, A. Sole-Daura, and Y. Wang, UMA-ASE: A FAIR-Aware Environment for Machine-Learned Interatomic Potential Workflows with UMA and ASE, ChemRxiv 2026, DOI: 10.26434/chemrxiv.15001161/v1

When preparing a manuscript, we also recommend stating the exact uma-ase version used in the calculations, together with the relevant UMA checkpoint and task settings, so the workflow can be reproduced accurately.

Download links:

• Paper landing page
• Version 1 PDF
• BibTeX entry

If the calculation outputs are deposited through ioChem-BD and a dataset DOI is assigned, cite both the deposited dataset DOI and the ioChem-BD platform paper. The recommended ioChem-BD references are:

• M. Álvarez-Moreno, C. de Graaf, N. López, F. Maseras, J. M. Poblet, and C. Bo, Managing the Computational Chemistry Big Data Problem: The ioChem-BD Platform, Journal of Chemical Information and Modeling 2015, 55, 95-103. DOI: 10.1021/ci500593j
• C. Bo, F. Maseras, and N. López, The role of computational results databases in accelerating the discovery of catalysts, Nature Catalysis 2018, 1, 809-810. DOI: 10.1038/s41929-018-0176-4

In short:

• cite the UMA-ASE paper when the software is central to published results
• report the exact software version and key model settings
• if results are uploaded to ioChem-BD, cite the dataset DOI and the ioChem-BD platform paper

Requirements

Validated with Python 3.12.

Core runtime:

• Python >= 3.10
• ASE >= 3.26
• numpy >= 2.2
• torch >= 2.6
• fairchem-core >= 2.10
• PyYAML >= 6.0
• requests >= 2.31
• ray[serve] >= 2.53.0

Optional features:

• Webapp and reports: flask, matplotlib, python-docx
• Analyze tab: pandas
• ioChem-BD tooling: importlib_resources, lxml, saxonche, networkx
• Local SMILES/InChI 3D generation: RDKit from conda-forge

Notes:

• On some Ubuntu installs, RDKit helper modules such as drawing or bond-inference extensions may be packaged separately. uma-ase now enables the core RDKit workflow as long as Chem and AllChem are available; optional features degrade gracefully when helper modules are missing.

Installation

Choose one of these installation scenarios.

Scenario 1: run directly from a source checkout

Use this when:

• you are developing UMA-ASE
• you do not want to install the package yet
• you are happy to start it with PYTHONPATH=src

Setup:

cd /path/to/UMA-ASE
pip install -r requirements.txt
conda install -c conda-forge rdkit

On Apple Silicon, you may also want:

export OMP_NUM_THREADS=1

Run:

export PYTHONPATH=src
python -m uma_ase.server

This is the workflow currently used in local development.

Scenario 2: install from source in editable mode

Use this when:

• you want the uma-ase and uma-ase-server commands
• you are still developing locally
• you want code changes in the checkout to be reflected immediately

Typical setup with a virtual environment:

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
pip install -e ".[server,iochem,analysis]"

On Apple Silicon, you may also want:

export OMP_NUM_THREADS=1

For SMILES/InChI 3D generation:

conda install -c conda-forge rdkit

Then run:

uma-ase-server

Scenario 3: install from a built wheel or tarball in dist/

Use this when:

• you want to deploy UMA-ASE on another machine
• you do not want to rely on a source checkout
• you want a repeatable installation from generated artifacts

This is the recommended deployment path.

Important:

• use one single python executable for the whole deployment
• do not check RDKit with one Python and then install or run UMA-ASE with another one
• after wheel installation, PYTHONPATH is no longer needed

In practice, this means:

1. find the Python executable that can import rdkit
2. use that exact executable path for installation
3. use that same executable path again to start the server

Example:

/usr/local/bin/python -c "from rdkit import Chem; from rdkit.Chem import AllChem; print('RDKit OK')"
/usr/local/bin/python -m pip install --user "uma-ase[server,iochem,analysis] @ file:///.../uma_ase-0.9.8-py3-none-any.whl"
/usr/local/bin/python -m uma_ase.server

If python3 does not see RDKit, but a conda-managed Python does, use the conda Python path instead.

On the source machine, build fresh artifacts:

cd /path/to/UMA-ASE
python -m pip install --upgrade build
python -m build

Then copy these files to the target machine:

• dist/uma_ase-0.9.8-py3-none-any.whl
• scripts/install_from_dist.sh

On the target machine, install from the wheel:

./scripts/install_from_dist.sh \
  --python "$(command -v python3)" \
  --checkpoint-dir "$HOME/uma-checkpoints" \
  --results-dir "$HOME/.uma_ase/results"

The installer:

• checks Python >= 3.10
• aborts if RDKit is not importable in that Python
• installs uma-ase with server, iochem, and analysis extras
• writes ~/.uma_ase/uma_ase_env.sh
• can append that env file to your shell startup file when --persist-shell is used

Load the generated shell snippet:

source ~/.uma_ase/uma_ase_env.sh

Then start UMA-ASE:

uma-ase-server

Alternative:

python -m uma_ase.server

Notes:

• default results directory: ~/.uma_ase/results
• default checkpoint directory written by the installer: ~/uma-checkpoints
• on macOS, the installer writes OMP_NUM_THREADS=1 unless it is already set
• if you install with --user, the installer also exports the Python user bin/ directory into PATH
• if the installer says the wheel is missing vendored ioChem stylesheet files, rebuild dist/ first
• for a long-running Linux service, you can install gunicorn and run gunicorn "uma_ase.server:create_app()" --bind 0.0.0.0:8000

Scenario 4: install from a package index

Use this when:

• UMA-ASE is already published to your package index
• you do not need a local dist/
• you want a normal pip install

Base install:

pip install uma-ase

Full webapp install:

pip install "uma-ase[server,iochem,analysis]"

Optional extras:

• server: webapp and reports
• iochem: ioChem-BD and CML/XSLT tooling
• analysis: Analyze-tab dependencies

If FairChem is not already present:

pip install fairchem-core

For SMILES/InChI 3D generation:

conda install -c conda-forge rdkit

Quick Start

CLI

export UMA_CUSTOM_CHECKPOINT_DIR=/abs/path/to/checkpoints
uma-ase -input geometry.xyz -chg 0 -spin 1 -run-type sp

From a source checkout without installing:

export PYTHONPATH=src
python -m uma_ase.cli -input geometry.xyz -chg 0 -spin 1 -run-type geoopt

Webapp

Installed package:

uma-ase-server

From source:

export PYTHONPATH=src
python -m uma_ase.server

Then open http://127.0.0.1:8000.

Configuration

UMA checkpoints

• UMA_CUSTOM_CHECKPOINT_DIR: directory containing UMA checkpoints such as uma-s-1p1.pt

Results and queue storage

• UMA_RESULTS_DIR: root directory for jobs, logs, bundles, and analysis reports
• job metadata is stored in ${UMA_RESULTS_DIR}/jobs.jsonl

ioChem-BD

• UMA_ASE_IOCHEM_REST_URL: default REST endpoint used for bare handles
• UMA_ASE_IOCHEM_CREATE_BASE: default Create REST endpoint
• UMA_ASE_IOCHEM_VERIFY: set to 0 to skip TLS verification
• UMA_ASE_IOCHEM_REST_FALLBACKS: fallback REST endpoints for bare handles only

Notes:

• when a full ioChem-BD URL is pasted in the webapp, the original host is preserved
• fallback hosts are used only for bare handles such as 100/6530

RDKit / PubChem

• UMA_ASE_MAX_CONFORMERS default: 2000
• UMA_ASE_RDKIT_MAX_BOND_ATOMS default: 200
• UMA_ASE_RDKIT_MAX_2D_ATOMS default: 250
• UMA_ASE_PUBCHEM_DEBUG=1: enable PubChem conformer diagnostics
• to diagnose RDKit availability on a running server, query:

curl http://127.0.0.1:8000/api/uma-ase/chem/diagnostics

• the diagnostics payload reports whether core RDKit imports, drawing support, and rdDetermineBonds are available, and whether a test 2D PNG can actually be rendered on that machine
• Generate 3D from SMILES or InChI prefers local RDKit rendering for the 2D sketch, but falls back to a PubChem PNG when the local drawing stack is incomplete

CLI

Main behavior

• -input accepts an ASE-readable geometry file or a webapp-generated .use recipe
• charge and spin can be read automatically from XYZ comment lines
• ts runs perform:
  1. an initial vibrational calculation
  2. a dimer search
  3. a final vibrational validation
• md must be run alone

Common options

• -input: geometry file or .use file
• -chg: molecular charge
• -spin: spin multiplicity
• -run-type: sp, geoopt, freqs, ts, md
• -iter: maximum optimization steps
• -grad: force threshold in eV/Å
• -optimizer: ASE optimizer for standard geometry optimization
• -mlff-chk: UMA checkpoint
• -mlff-task: UMA task
• -temp / --temperature: vibrational-analysis temperature in K
• -press / --pressure: vibrational-analysis pressure in Pa
• -cpu: force CPU execution
• -visualize: open the optimized structure in ASE's viewer

TS options

• --ts-displacement
• --ts-dimer-separation
• --ts-trial-step
• --ts-max-step
• --ts-max-rot

MD options

Selected groups:

• Core: --md-engine, --md-steps, --md-timestep-fs, --md-temperature, --md-traj-interval, --md-log-interval
• Thermostats: --md-friction, --md-relax-fs, --md-nose-*, --md-bussi-taut-fs, --md-andersen-prob
• Barostats / NPT: --md-pressure-bar, --md-compressibility-bar, --md-barostat-relax-fs, --md-mtk-*, --md-baoab-*, --md-melchionna-*
• PBC / cell: --md-use-pbc, --md-cell-a/b/c, --md-cell-alpha/beta/gamma
• Restart: --md-restart-traj

Use uma-ase -h for the full reference.

XYZ charge / spin metadata

uma-ase reads charge and spin from XYZ comment lines when present, for example:

199
https://iochem-bd.iciq.es/browse/handle/100/69781 charge=-1 spin=2
O  0.0 0.0 0.0
H  0.0 0.0 0.96
H  0.0 0.75 -0.48

Legacy signed-integer comment patterns are also accepted.

.use recipe files

Every webapp job writes a *.use JSON recipe containing:

• full input geometry payload (filename, format, original saved path, and embedded content)
• run type / run sequence
• method metadata (method = MLIP-UMA, with the selected checkpoint stored as checkpoint)
• all run-type-specific parameters needed to reproduce the job
• generator metadata (uma-ase version and timestamp)

The webapp validates the parameter set before writing the .use file, so a recipe is only saved when the required fields for that run type are present. Older .use files that still use mlff_chk remain readable for backward compatibility.

Reuse from CLI:

uma-ase -input /path/to/job.use

Reuse from the webapp:

1. Open Load/Edit
2. Click Load .use
3. Review the restored parameters
4. Go to Calculate and launch again

Logs and timing

Each run writes a Run metadata block with the minimal information needed to reproduce the job.

• sp and geoopt omit irrelevant thermochemistry metadata
• freqs and ts include temperature and pressure
• md includes only the active engine and active PBC / cell settings
• freqs and ts explicitly re-symmetrize the finite-difference Hessian with (H + H^T) / 2 and project rigid translations/rotations before diagonalization
• -FREQS-modes.zip now includes every vibrational mode trajectory, including near-zero modes, and uses normalized fixed-amplitude mode animations for stable visualization

Timing policy:

• log blocks use a consistent sectioned layout for CLI arguments, atomic-system summary, run metadata, geometry tables, calculator setup, and final result summaries
• each CLI log starts with a UMA-ASE version banner before the argument block
• main run headers now use the same full-form naming style across Single Point, Geometry Optimization, Vibrational Analysis, Transition State Search, and Molecular Dynamics
• each context switch now inserts two blank lines before the next ===== or ----- banner block for easier visual scanning
• Timing [...] banner lines now print as closed separator blocks above and below the timing line, and bonding-energy sections include a short note describing the single-atom UMA reference setup
• per-phase timing reports wall time
• log events with timestamps are emitted in a highlighted banner style for run starts, completions, and major phase transitions
• setup phases now include explicit ASE-load timing, including geometry loading, UMA checkpoint loading, calculator setup, and MD restart-trajectory loading when applicable
• atomic reference energies now reuse the already loaded UMA predictor within each run instead of reloading the checkpoint per atom type
• single-atom reference calculations now use the neutral-atom ground-state multiplicity instead of forcing multiplicity 1
• the packaged multiplicity table lives in src/uma_ase/data/ground_state_atomic_multiplicities.json, and the bonding-energy log section prints the data provenance used for those references
• the final Run total line reports both CPU and wall time

Webapp

The webapp combines geometry intake, submission, visualization, analysis, queue management, and ioChem-BD integration.

Load/Edit

Use this tab to import or build the atomic system.

Supported sources:

• local files such as .xyz, .pdb, .mol, .cif
• ioChem-BD handles and full URLs
• PubChem CID or name
• PDBe entry ID
• SMILES / InChI via RDKit

Highlights:

• Atomic System Details summarizes charge, multiplicity, formula, atom count, 2D sketch, IUPAC name, and conformer state
• successful fetch/generation keeps the active source input and clears the other source fields
• loading or creating a new structure clears the Visualize tab state before the new system is previewed
• PubChem also copies the returned SMILES into the SMILES box
• non-periodic CML is kept as CML for viewing
• periodic CML is converted to CIF before loading into JSmol
• static jobs convert CML to XYZ for ASE; MD uses CIF if periodic, otherwise XYZ
• charge and multiplicity extracted from ioChem-BD CML are propagated to Calculate and MD
• RDKit-derived 2D sketch, IUPAC name, and inferred SMILES are skipped for systems larger than 150 atoms

Conformers:

• generate from SMILES or from the current structure
• navigate with the isomer / conformer controls above the viewer
• UFF energies are shown when available
• generation can be stopped from the UI

Editing:

• select an atom in JSmol
• change its element in the Edit card
• use DO, UNDO, and REDO

Periodic systems:

• MD PBC is off by default
• MD cell fields are filled only when the structure contains periodic data
• stale PBC values are cleared when a new non-periodic structure is loaded

Calculate

This tab combines static calculations, multi-job submission, and MD.

Modes:

• Single Job
• Multi Jobs
• Molecular Dynamics

Common layout:

• selected system filename shown in the System card
• charge and multiplicity on the next row
• method selector, UMA checkpoint, and UMA task

Static calculations:

• supports sp, geoopt, freqs, and ts
• the output panel streams the full log
• Download Results bundles all generated files

Vibrational settings:

• shown as Vibrational Analysis Settings
• webapp defaults: 300 K and 1 atm
• backend and CLI keep pressure in Pascals

TS behavior:

• TS reuses the same visible controls as GeoOpt + Frequencies
• optimizer, force threshold, max GeoOpt iterations, temperature, and pressure remain visible
• the optimizer is forced to Dimer algorithm while TS is selected

Multi Jobs:

• each selected input file becomes its own queued job
• results are bundled into a single ZIP

Molecular Dynamics:

• exposes ASE MD engines and their thermostat / barostat controls
• supports editable PBC and cell parameters
• live log output is shown in the MD output panel
• periodic inputs are passed to ASE as CIF

Visualize

Use this tab to inspect completed or running results.

It provides:

• the main JSmol structure viewer
• a trajectory viewer for .traj, .xyz, and live MD capture
• a vibrational-mode loader for *-FREQS-modes.zip plus the matching log
• a completed-job loader through .use and .log

Important behavior:

• freqs and TS runs auto-populate the vibrational-mode controls after completion
• completed periodic MD jobs keep their cell during playback
• completed periodic MD playback is recentered to the frame center of mass
• opening a job from My Jobs preloads the relevant visualization data

Analyze

The Analyze tab runs post-processing workflows directly in the browser.

Available analyses:

• RMSD Report
• XYZ Pair Analysis
• Simple Geometry Analysis

Shared workflow:

• choose an analysis card
• load the required file or folder directly in that card
• use the shared action bar to run and download the selected format
• monitor the shared progress bar during execution

Outputs:

• RMSD Report: plot plus HTML summary/tables
• XYZ Pair Analysis: 3D scatter, RMSD-vs-charge plot, and tables
• Simple Geometry Analysis: summary tables plus histogram

My Jobs

This tab shows the persistent server-side queue and completed-job history.

Features:

• queued / running / completed / interrupted states
• submission time and elapsed runtime
• per-job actions for view, rename, restart MD, kill, download, delete
• bulk delete and download
• job-count and storage-usage badges
• retention countdown indicator per job

View behavior:

• SP: load the log into Calculate and stay there
• Freqs: load the log into Calculate and the normal-mode data into Visualize
• GeoOpt / TS: load the log into Calculate and the optimized trajectory into Visualize
• MD: load the log into both Calculate and MD output, and open Visualize with the trajectory
• running jobs opened from My Jobs switch to Visualize and update only the tracking / visualization widgets until completion
• when a tracked Freqs job reaches completed, its final log is also restored in Calculate
• when a running Freqs job auto-restores the initial structure for visualization, the Calculate log is preserved instead of being replaced by a generic Loaded ... message
• before a completed GeoOpt / Freqs / TS result is auto-loaded, the existing Visualize state is cleared so stale trajectory or mode controls do not persist
• loading a completed MD trajectory directly in Visualize also clears the tab state first, so stale trajectory or mode controls do not persist

Queue history persists across server restarts through jobs.jsonl.

ioChem-BD

This tab provides browsing and upload support for ioChem-BD.

Workflow:

1. Log in to obtain a JWT token
2. Browse your tree or shared trees
3. Select a destination path
4. Upload results from a completed UMA job

Behavior:

• browse and upload cards stay inactive until a token is available
• upload activates only after a destination path is loaded
• one tree applet is used for both personal and shared content
• the filter below the applet filters the currently displayed tree
• when a full ioChem-BD URL is pasted, the original host is preserved

Recommended import flow:

1. Open the desired calculation in ioChem-BD
2. Copy the handle URL
3. Return to Load/Edit
4. Paste the handle and load it

About

Displays this README inside the application. The left-side section menu is generated dynamically from the README headings, opens first-level sections by default, keeps deeper subsections folded until expanded, tracks the current section while you scroll, and can be filtered live. The LICENSE link opens the GPL text in an in-app overlay.

Output Files

Generated files depend on the workflow that actually ran.

Static runs

• input.xyz
• <stem>.use
• <stem>-<LABELS>.log
• <stem>-<LABELS>.xml
• -geoopt-<LABELS>.xyz and -<LABELS>.traj for GeoOpt / TS
• <stem>-FREQS-modes.zip for Freqs / TS

MD runs

• <stem>-MD.log
• <stem>-MD.xml
• <stem>.use
• input.cif when periodic, otherwise input.xyz
• <stem>-MD.traj
• <stem>-MD.xyz
• <stem>-MD-final.xyz

Multi Jobs

• one subfolder per input file
• each subfolder contains the same outputs as the corresponding single workflow

Package Layout

src/uma_ase/
├── __init__.py
├── __main__.py
├── cli.py
├── server.py
├── utils.py
├── workflows.py
└── static/uma-ase.html

Development

Typical workflow:

1. create a virtual environment
2. install in editable mode with the needed extras
3. run tests or manual checks
4. build distributions

Build:

python -m build

Upload:

• GitLab Package Registry

  python -m twine upload \
    --repository-url "https://gitlab.com/api/v4/projects/<project-id>/packages/pypi" \
    dist/*
  

• PyPI

  python -m twine upload dist/*
  

License

Copyright (c) This code was generated using OpenAI Codex agents following the guidance of humans. C. Bo, ICIQ, 2025-2026. Distributed under the GNU General Public License v3.0. See LICENSE.