RedoxFlow: Agentic workflow for preparing simulations in high-throughput redox-potential screening

RedoxFlow generates candidate organic molecules, proposes reduced products, and auto-prepares simulation inputs to compute thermodynamic (Nernstian) redox potentials in aqueous phase via the Computational Hydrogen Electrode (CHE) method [1].

---

Why Redox Potentials?

Redox potential quantifies how readily a molecule is oxidized or reduced. Mapping E across chemical space is a fast screener of thermodynamic driving force (via ΔG = −zFE): it lets you quickly rank candidates, estimate feasible cell voltages, and check compatibility with solvent/electrolyte stability windows and pH before doing any heavy kinetic/mechanistic work. For Proton-coupled electron-transfer (PCET) steps, CHE also exposes the pH-dependence (Nernst slope), enabling screening across operating conditions by adjusting calculated redox potentials for pH after-the-fact.

Redox potentials are especially useful in:

• Batteries / Flow batteries: target voltage windows (∆E ≈ E_cathode − E_anode) while respecting solvent/electrolyte stability.
• Bio/Photoelectrochemistry: choose mediators/analytes whose E minimizes O₂/H₂ interference and matches desired driving force.
• Environmental electrochemistry: assess spontaneity/selectivity along the redox ladder for contaminant transformations.
• Redox-swing separations / CO₂ capture: tune E to set binding–unbinding windows and energy efficiency.

CHE computes redox potentials from a thermodynamic cycle. RedoxFlow automates input preparation and post-processing—extracting energies and computing (E=-ΔG/(zF)) from completed runs. It does not launch or manage simulations. Pair it with your scheduler or workflow engine (e.g., Slurm, Nextflow, Snakemake) to run jobs at scale; this project’s focus is robust preparation and calculation, enabling high-throughput screening when integrated into a broader execution pipeline.

---

TL;DR

• Generates reactants (de novo) with a lightweight foundation model (ibm-research/GP-MoLFormer-Uniq [2])
  • Unrealistic moieties/ring strained systems auto-filtered out
  • Generated molecules with non-CNOF atoms are altered within their same periodic group (e.g., S → O) to maximize generative success
  • Sanitized molecules to ensure proper basic connectivity/valency
• Predicts reduced products (transparent rules based off functional groups, proof-of-concept)
  • Live memory during agent use is auto-created at RedoxFlow/memory but we share example memory from a previous run at RedoxFlow/example_memory
• Does conformer search and picks lowest-energy structures
• Writes simulation scripts for both states and computes (E = -ΔG/(zF)) for completed simulations

The agent is loaded locally from within the repo and thus is deployable on local machines or HPC for script generation provided that the environment is installed.

---

Install

Tested on HPC (PSC Bridges-2, Linux-based)

Tested locally (OS: Ubuntu 22.04.5 LTS, GPU: GeForce RTX 3090)

git clone <this-repo> RedoxFlow
cd RedoxFlow
mamba env create -p ../redoxflow -f env/redoxflow.yml
conda activate ../redoxflow

The agent generates simulation inputs and calculates redox potential from successful simulations but does not launch/manage simulations. We forego doing this due to resource costs but envision it as future steps for the project.

For our test cases RedoxFlow/redox_calculation_test, we run the simulations via an NWChem docker image and provide reproducible steps below.

Installing NWChem Docker Image

# 1) Make sure Docker Engine is installed (skip if you already have it)
#    Install guide: https://docs.docker.com/engine/install/

# 2) Pull the current MPI-enabled NWChem image from GHCR
docker pull ghcr.io/nwchemgit/nwchem-dev.mpi-pr:latest

# 3) Confirm it’s on your machine
docker images | grep nwchem

Agent Demonstration

Generate Reactants & Auto-prepare Scripts & Calculate Redox Potential

RedoxFlow/scripts/agent_demonstration.ipynb

Running NWChem Simulations (to validate script preparation)

How to run a prepared simulation by the agent (32 MPI ranks were used in this case for a 32-core CPU)

Example Script Folder:

/path/to/RedoxFlow/redox_calculation_test/reactants/react_2

# 1) Set your job folder (note the quotes for the space)
HOSTDIR="/path/to/RedoxFlow/redox_calculation_test/reactants/react_2"

# 2) Run a single job (32 MPI ranks; 1 OpenMP thread each)
docker run --rm --shm-size=1g \
  -e MYNPROC=32 -e OMP_NUM_THREADS=1 \
  -v "$HOSTDIR":"$HOSTDIR" \
  -w "$HOSTDIR" \
  ghcr.io/nwchemgit/nwchem-dev.mpi-pr:latest \
  react_2.nw > react_2.out 2>&1

Or, for the equivalent product,

Example Script Folder:

/path/to/RedoxFlow/redox_calculation_test/products/prod_2

# 1) Set your job folder (ABSOLUTE FILEPATH)
HOSTDIR="/path/to/RedoxFlow/redox_calculation_test/products/prod_2"

# 2) Run a single job (32 MPI ranks; 1 OpenMP thread each)
docker run --rm --shm-size=1g \
  -e MYNPROC=32 -e OMP_NUM_THREADS=1 \
  -v "$HOSTDIR":"$HOSTDIR" \
  -w "$HOSTDIR" \
  ghcr.io/nwchemgit/nwchem-dev.mpi-pr:latest \
  prod_2.nw > prod_2.out 2>&1

---

Proof-of-Concept Restrictions

To show that the agentic workflow works start-to-finish, we restrict our agent to generate molecules and prepare simulation scripts for:

• CNOF-containing molecules
• A model reduction reaction A + xH^+ + ye^- → B
  • No degradation products considered
  • By nature of the model reaction, we only consider PCET-reactions.
• PBE Functional
• def2-SV(P) basis set
• Pure water (dielectric constant ~ 78.4)
• Room temperature

For a round-trip demonstration of redox potential calculation with the agent/embedded class interface, we:

• Generate 3 simulations (3 reactants/3 products) and thus auto-extract variables and auto-calculate 3 redox potentials (in demonstration notebook)
  • /path/to/RedoxFlow/redox_calculation_test
• An intentionally incomplete simulation to show error tracking to motivate extended automation pipelines with simulation management (in demonstration notebook)
  • /path/to/RedoxFlow/redox_calculation_test/react_3, /path/to/RedoxFlow/redox_calculation_test/prod_3

---

Envisioned Improvements on Proof-of-Concept Automation

The agent workflow provided is a minimum proof-of-concept for simulation preparation start-to-finish and redox potential calculation start-to-finish. The automation premise is extendable based on what we see as future steps to the work:

Automation Improvements

• Pair the agent with an en-masse job submission workflow (SLURM, Nextflow, Snakemake, etc.)
• Adapt the .csv-based memory to a proper relational database (e.g., PostgreSQL, SQLite)
• Expand the script preparation method (RedoxFlow.prepare_scripts()) to include fine-tuned DFT simulation settings (e.g., basis set, functional, dielectric constant, temperature)
  • We currently restrict these settings to tractable, fixed values deliberately for proof-of-concept
• Pair the script preparation method (RedoxFlow.prepare_scripts()) with a natural language prompt interface
  • e.g., "Generate me 100 reactant candidates using the PBE functional and 6-311g* basis set at 200 Kelvin."

Chemistry Improvements

• Expand reaction rules or replace the reaction rule proof-of-concept with a partial charge predictive model to auto-identify likely reduction sites (e.g., Gasteiger partial charge assignment)
• Allow the user to specify their own blacklisted moieties to restrict chemically unrealistic reactants
  • We currently set this blacklist minimally inside redoxflow.py as proof-of-concept but could easily expose this to the interface for the researcher.
• Expand the subset of atoms (e.g., include Cl, P, S)
  • We restrict the chemistry to CNOF molecules for proof-of-concept demonstration.

---