Research Scientist, Machine Learning

Who we are

Arena Physica is on a mission to accelerate hardware innovation that powers human progress. Our name is inspired by Theodore Roosevelt's 'Citizenship in a Republic' speech. To us, entering the Arena means committing fully and accepting the risk of failure in pursuit of an audacious, worthy cause. We believe the future belongs to those brave enough to build it.

Our team of 50 combines AI engineering and applied physics expertise with deep experience in enterprise deployments. We're headquartered in NYC with presences in San Francisco and Los Angeles, backed by ~$90M from Initialized, Founders Fund, Goldcrest Capital, Fifth Down Capital, and Shield Capital.

If you're ready to do the most important work of your career, join us in the Arena.

What we do

At Arena Physica, we're building electromagnetic superintelligence. Our AI platform Atlas operationalizes physics-grounded intelligence to verify, debug, and optimize hardware across its lifecycle. Atlas is already trusted globally by the world's most advanced hardware companies, including AMD, Anduril, and Bausch & Lomb, for applications across R&D, integration testing, production assembly, and field repair.

About the role

As a Research Scientist — Machine Learning, you will set the technical direction for a core slice of Arena's electromagnetic foundation model: model architecture, training and evaluation pipelines, data strategy, and the publications that demonstrate it works. You'll work directly with our existing and fast-growing research team and partner with senior leadership on the proof-points roadmap.

This is a truly unique opportunity at the intersection of a renewed industrial and political focus on hardware and the emergence of AI for Physics. At Arena Physica you will work with diverse expert teams across Software, Hardware, and ML. If you are ambitious, drawn to big bets, and truly excited about transforming how we apply physics to hardware development this role is for you.

How you will contribute

* Designing our Model Architecture - Design and prototype model architectures grounded in EM physics and applied math — neural operators (FNO, AFNO, deformation-FNO, U-FNO), Transolver-class architectures, foundation-model backbones, and world-model-style joint-embedding predictive architectures, or geometry focused encoders (GeoPT, DeepSDF, point-cloud / mesh transformers). - Make principled choices based on physics priors: PINN losses, structural inductive biases, conservation enforcement, passivity / reciprocity constraints, and exterior-calculus-style discretization.

* Fast Experimentation - Build and run end-to-end training and evaluation pipelines: dataloading, distributed training, ablations, scaling experiments, and large-model orchestration. - Drive experimental discipline — pre-registered hypotheses, paired comparisons, parameter-matched runs, confidence bands on results.

* Expand the Data Strategy - Partner with the team on training-corpus composition, augmentation, curriculum design, and physics-informed multipliers to drive learning in and out of distribution. - Specify coverage targets and the simulation campaigns required to achieve them; close the loop with Arena's solver-farm, sim infrastructure, and lab measurements.

* Translate Research into Atlas Deployments - Partner with the Atlas product and platform engineering teams to land research artifacts in production deployments across our customer base. - Define inference performance, accuracy, and uncertainty requirements that real-world use cases need; close the loop from production data back into research direction.

* Author Research Publications - Author papers at top venues (NeurIPS, ICML, ICLR, IMS) that reinforce Arena's research footprint and externally validate the foundation-model approach. - Own at least one substantive publication per year as first author.

* Drive External Collaborations - Develop and support research collaborations with academic labs and industry partners. - Represent Arena externally at conferences, workshops, and customer engagements.

You have

* PhD or equivalent research track in Electrical Engineering, Applied Physics, Applied Mathematics, or Computer Science.

* Deep background in electromagnetic physics or in the applied mathematics of PDE solvers — ideally directly in EM.

* Demonstrated focus on foundation models, neural operators, world models, Transolver-class architectures, or related operator-style ML.

* A published track record at top ML or computational-physics venues — first-author or senior-author papers.

* Strong intuition for the mathematics behind EM solvers (FEM / FDTD / MoM / eigenmode / Green's functions) and how that intuition translates into architectural choices.

* Proficiency in Python and the common ML framework + tooling stack — frameworks such as PyTorch and JAX, distributed-training stacks (e.g. FSDP, DeepSpeed), experiment tracking (e.g. MLflow, Weights & Biases), and performance tooling (e.g. torch.compile, XLA, CUDA); comfort writing C / C++ when needed for performance-critical paths.

* Track record of shipping experiments end-to-end — not just architecting in slideware, but actually running pipelines at meaningful scale.

* [Preferred] Prior work integrating physics priors directly into model architectures (e.g., PINN losses, exterior-calculus-style discretizations, scale-invariance constraints).

* [Preferred] Direct experience with EM simulators such as ADS, HFSS, MATLAB EM, or comparable solvers.

* [Preferred] Experience with HPC ecosystems.

Benefits & Perks Include:

* 100% of the monthly premiums covered with Aetna medical vision, and dental insurance for you and your dependents

* 401(k) Retirement Plan

* Unlimited PTO

* Lunch every day from local restaurants via Sharebite

* Relocation support provided