Machine Learning Scientist - Molecular Force Fields

SandboxAQ; United States; Remote; USD 120,000 - 190,000; Full time

Job summary

SandboxAQ is seeking a mid-level Machine Learning Scientist to own ML force fields for PFAS substitution within the ChemSim group. The role involves designing high-throughput quantum chemistry and MD campaigns, training ML interatomic potentials, and extending them into PFAS-related chemistries.

You will build reproducible pipelines, validate models across large chemical spaces, and deploy trained models on SandboxAQ tooling so scientists can reuse workflows and data in production environments.

Qualifications

• PhD in Computational Chemistry, Materials Science, Physics, Machine Learning, or a related field with deep specialization in ML interatomic potentials.
• Experience deploying ML force fields and shipping them for use inside a company or publicly.
• Hands-on data generation for first-principles training at scale with quantum chemistry codes.
• Fluency with descriptor-based, equivariant, and message-passing potentials; know strengths and limits.
• Strong Python programming and modern ML framework experience, including GPU training.
• Experience integrating learned potentials with production MD engines and extracting observables.
• Experience building automated workflows and curating large scientific datasets with provenance.
• Proven validation discipline and handling of out-of-domain cases; not only best-case benchmarks.
• Experience in cloud and HPC environments.
• Hands-on experience with fluorine-containing chemistry, fluoropolymers, or PFAS-adjacent systems.

Responsibilities

• Own the training data: design high-throughput physics campaigns and set theory level, basis, convergence, and sampling criteria.
• Build data generation infrastructure: automate, restartable workflows across cloud/HPC; curate datasets.
• Train and own MLFF stack: choose architectures, train, validate, and deploy potentials across PFAS space.
• Deliver property predictions: develop MLFF-driven protocols for protonation, diffusivity, viscosity, and thermal conductivity.
• Extend potentials to PFAS substitution chemistries: resin polymerisation and plasma surface chemistry.
• Run production campaigns and deploy models: scale predictions and publish documentation for reuse.

Skills

Python programming

Machine learning

Molecular simulation

High performance computing

Cloud computing

Education

PhD in Computational Chemistry/Materials Science/Physics/ML

Tools

VASP

Quantum ESPRESSO

CP2K

Gaussian

LAMMPS

GROMACS

OpenMM

ASE

PyTorch

JAX

Job description

Machine Learning Scientist - Molecular Force Fields at SandboxAQ, United States (remote). A mid-level full time position in the ChemSim team, listed under software engineering.

The Opportunity

Introduction to the team: This role is for the PFAS team which sits within SandboxAQ's Chemical Simulation (ChemSim) group. Our mission is to develop PFAS-lean or PFAS-free substitutes and formulations for semiconductor process and fab materials that meet both performance specifications and environmental and safety requirements in complex semiconductor manufacturing environments. We combine generative ML, physics based simulation (e.g. DFT and molecular dynamics), Large Quantitative Models for property prediction, and multi-scale modeling with a partner driven experimental validation loop. You will work alongside industrial co-development partners to move candidate molecules from prediction to qualified use.

Introduction to the role: The PFAS team is hiring Machine Learning Scientists to own machine learned force fields (MLFFs) for the PFAS substitution programme, including generating the first principles data they are trained on. Machine learned potentials are the load bearing element of our discovery pipeline: the properties that decide whether a candidate molecule can replace a PFAS incumbent — among them reactivity, thermal stability, and the diffusivity through polymer matrices — are ones that the physics based simulations can compute accurately but far too slowly to screen a library, and that classical force fields compute quickly but may not be accurate enough to trust. In this role you will: (1) design and run the high-throughput quantum chemistry and molecular dynamics campaigns that produce our training data, and own the infrastructure that makes them reproducible; (2) train, validate, and deploy the potentials that data supports, and identify where each model's domain of validity ends; and (3) extend machine learned potentials into the chemistry that PFAS substitution requires. See how SandboxAQ is helping build America's semiconductor supply chain from the materials up

Responsibilities

• Own the training data. Design and run high-throughput physics based simulation campaigns that produce labelled training data for the PFAS chemical space, making the calls that determine whether the data is fit to train on — level of theory, basis set, convergence criteria, sampling strategy, and error budget.
• Build the data generation infrastructure. Stand up automated, restartable, provenance tracked workflows across cloud and HPC by collaborating with the engineering team; curate and version the resulting datasets, and combine internally generated first-principles data with relevant public molecular and materials datasets under consistent conventions.
• Train and own the MLFF stack for the PFAS vertical. Select architectures, train, validate, and deploy machine learned interatomic potentials across the chemical space of PFAS substitution demands. Fine tuning universal or foundation potentials, or training from scratch when appropriate.
• Deliver the property predictions the screening pipeline runs on. Build and validate MLFF-driven protocols for the properties our pipelines need — for example protonation and deprotonation energetics underpinning pKa, ion diffusivity in glassy polymer matrices, and Green-Kubo viscosity and thermal conductivity — each benchmarked against reference data or experiment.
• Push the potentials into the process chemistries that matter. Extend coverage to resin phase polymerisation and plasma surface etch chemistry, moving beyond single molecule, gas phase behaviour into the condensed phase regime our applications actually occupy.
• Run production campaigns and deploy models on our platform. Execute at scale prediction campaigns across full candidate libraries, and land trained models, datasets, and reproducible training pipelines in SandboxAQ tooling so that other scientists can run them from written documentation.

Requirements

• PhD in Computational Chemistry, Materials Science, Physics, Machine Learning, or a related field, with deep specialization in machine learned interatomic potentials or atomistic simulation.
• Demonstrated depth developing and deploying machine learned force fields — including at least one potential that you trained, validated, and put into the hands of other people who now rely on it, whether inside a company or as a public release. We are interested in what you have built and shipped rather than in years served, and this posting is open across experience levels.
• Hands-on experience generating first principles training data at throughput with production quantum chemistry and electronic structure codes (e.g. VASP, Quantum ESPRESSO, CP2K, ORCA, Psi4, or Gaussian), including the judgment to set and defend level of theory, basis set, and convergence criteria.
• Demonstrated fluency with descriptor based, modern equivariant, and message-passing potential architectures (e.g. MACE, NequIP, Allegro, TensorNet, DeePMD, ANI-class, or FairChem/UMA-family models) — including what each is good for and where each fails.
• Strong programming skills in Python and hands on experience with modern ML frameworks (e.g. PyTorch, JAX), including training at scale on GPU clusters.
• Practical experience coupling learned potentials into production molecular dynamics engines (e.g. LAMMPS, GROMACS, OpenMM, ASE) and extracting converged physical observables from the resulting trajectories.
• Experience building automated computational workflows and curating large scientific datasets — provenance, versioning, and reproducibility.
• Demonstrated validation discipline: a track record of characterising model error and out-of-domain failure rather than reporting best case benchmark numbers.
• Experience working in cloud and high performance computing environments.
• Hands-on experience with fluorine containing chemistry, fluoropolymers, or PFAS-adjacent molecular systems.