The posting
ABOUT THE COMPANY
NeurAtom develops neural-network surrogate models for reactor-physics simulation. Full Monte Carlo neutron-transport calculations — the reference method for reactor core design and fuel optimisation — take hours on high-performance computing clusters. Our surrogates reproduce the quantities of interest (spatial power distribution, k-eff, reactivity) at sub-millisecond evaluation, validated against Monte Carlo ground truth. This makes design-space exploration and fuel-loading optimisation, currently constrained by simulation cost, tractable as interactive problems.
The context: a substantial expansion of nuclear construction is underway across Europe, with a large number of SMR and Gen-IV designs entering the design phase. Each new core requires extensive neutronics analysis. The computational methods in current use are either high-fidelity and slow (Serpent, OpenMC, MCNP) or fast and approximate; our work aims to reduce that trade-off. NeurAtom is a deep-tech spin-out from KTH, currently in the KTH Innovation Launch Programme.
ACCOMPLISHMENTS
NeurAtom is at an early stage. Current status, stated plainly:
- Two integrated prototypes.
- Validation against a trusted reference. Surrogates validated against Serpent 2 Monte Carlo (JEFF-3.1.1) on the SUNRISE-LFR lead-cooled fast-reactor geometry, using approximately 120,000 simulations on the Dardel national supercomputer.
- Known limits. All validation to date is on a single reactor geometry in an academic setting. The methodology has not yet been applied to industrial proprietary data — closing that gap is the substance of this role.
- Institutional standing. Accepted into the KTH Innovation Launch Programme (Batch 23), with access to premises, coaching, network, and proof-of-concept funding.
THE ROLE
The methodology performs well on the geometry it was developed on. The central objective is to make it transfer reliably to reactor designs and datasets it has not seen, and to develop it into a maintainable product. This position owns the surrogate-modelling methodology end to end. Principal areas of responsibility:
- Training-data strategy and adaptive sampling. Design how the input design space is sampled, and build the active-learning loop that selects which simulations to run next. Monte Carlo ground truth is expensive; this determines the cost of onboarding each new reactor geometry and is the primary lever on whether a reactor-agnostic product is economically viable.
- Data conditioning and validation. Distribution diagnostics, handling of heavy-tailed targets, normalisation choices and their failure modes, and stratified and adversarial validation.
- Validation envelope and uncertainty quantification. Characterise where predictions are reliable and where they are not, including out-of-distribution behaviour, and attach calibrated uncertainty to predictions. In a safety-relevant, licensing-adjacent context this is a requirement, not an enhancement.
- Optimisation over surrogates. Own the optimiser that searches loading patterns, including the interaction between surrogate error and an optimiser that will exploit regions where the surrogate is least accurate.
- Architecture selection and benchmarking. Benchmark architectures where the differences are consequential — sparse, high-gradient regions of the design space — without over-investing where models converge to similar performance.
- Reproducibility. Experiment tracking, model and dataset versioning, and regression testing, so that the validation record is auditable rather than reconstructed from memory.
You would be the sole owner of this workstream, working directly with the founders. There is no ML team above you and no established process to inherit.
The role calls for sound judgement on scientific data more than for coverage of a long tool list.
Required:
- Strong PyTorch, including reproducible training pipelines and models deployed beyond notebooks.
- Experience training models on simulation-generated data (Monte Carlo, CFD, FEA, or comparable), where ground truth is computationally expensive and spatially or temporally correlated. The domain need not be nuclear; this is the most relevant single qualification.
- Sound numerical and statistical judgement on scientific data — able to interpret residuals and error distributions and reason about their causes, and aware that aggregate metrics can conceal locally significant errors.
- A rigorous approach to validation.
- Ability to work independently as the sole owner of a workstream.
- Fluency in English (Swedish is a plus).
Meriting — depth in one or two of these is more valuable than familiarity with all:
- Active learning and adaptive sampling; uncertainty quantification. These are the highest-priority areas for the role.
- Transfer learning across problem instances; black-box or combinatorial optimisation.
- HPC workflows (SLURM, batch orchestration).
- Experiment tracking and model/dataset versioning.
- Reactor physics or neutron transport. Not required — the domain can be learned; the numerical judgement is harder to acquire.
Education: MSc or PhD in a computational discipline (computational science, applied mathematics, physics, engineering, or ML for physical systems), or equivalent demonstrated experience.
WHAT WE OFFER
- Equity as a founding team member, with reverse vesting on incorporation.
- Full technical ownership of the surrogate-modelling work, with the corresponding autonomy over methodology and tooling.
- A well-defined and technically demanding problem: expensive, correlated, simulation-generated data measured against a clear validation reference.
- KTH Innovation Launch Programme: premises at KTH campus, business coaching, network, and access to proof-of-concept funding.
- The role is not currently salaried. NeurAtom is pre-incorporation and pre-funding; compensation at this stage is equity.
