Overview

Physical Superintelligence is a startup with roots at Google, NVIDIA, Harvard, Meta, MIT, Princeton, Oxford, Johns Hopkins, Cambridge, and the Perimeter Institute building AI systems to discover new physics at scale. We are seeking applied physicists, engineers from the traditional disciplines (mechanical, electrical, aerospace, chemical), and computational scientists to put AI to work on real engineering systems: thermal and fluid systems, electrical power distribution, and the coupled models that tie them together. Models that beat the solvers they replace, and AI that designs rather than only evaluates. Applied AI at PSI is applied physics plus AI.

Our mission is to discover and commercialize transformative physics breakthroughs at scale with artificial superintelligence, safely, verifiably, and for broad public benefit.

The last century's golden age of physics gave us transistors, lasers, and nuclear energy. We believe artificial superintelligence will unlock the next one. We're creating the infrastructure to industrialize scientific discovery and usher in this new era.

We have one product: new physics, at scale.

Role and Responsibilities

Model real physical and engineering systems and validate the models against solvers and measured data.

Use learned models and simulators together: validate them against physics and measured data, decide where solver fidelity is required and where a learned model is enough, and use them to search a design space.

Go deep in one lane: fluids and heat transfer; thermodynamics and energy conversion; electrical and power systems; electrochemistry; solid mechanics and materials; transport and environmental physics; chemical and process engineering.

Ship to customers: each engagement solves their actual problem, ends with something they can run, and leaves behind a capability we reuse.

What We're Looking For

Three or more years building physics-based models of real engineering systems: simulations, reduced-order models, or first-principles analyses you set up yourself. You know which assumptions, boundary conditions, and mesh or discretization choices drive the answer, you have checked your results against experiments or measured data, and you can say what you did when they disagreed.

Three or more years applying scientific machine learning to real physics or engineering problems, not synthetic benchmarks. You have used, validated and deployed learned models such as surrogates, neural operators, or physics-informed networks, and you know how they behave out of distribution: where they extrapolate badly, how you detected it, and when you fell back to the solver.

Depth in one of the following engineering domains, with its codes and the standards it works to. You know where the models lie.

Fluid dynamics and heat transfer: CFD, conjugate heat transfer, multiphase and two-phase flow, natural and forced convection.

Thermodynamics and energy conversion: refrigeration and power cycles, heat exchangers, combustion, engines and turbines, HVAC.

Electrical and power systems: power flow, interconnection and protection, power electronics and converter dynamics, electromagnetics.

Multiphysics, coupled with digital twins/industrial applications expertise

Nice to Have

A PhD, or an equivalent research record, in engineering, applied physics, computational science, or machine learning for science.

A model you put into a system other people depend on: behind an API, versioned, with tests that catch a wrong number.

Training runs you designed and debugged yourself in PyTorch or JAX; neural operators, physics-informed networks, or uncertainty quantification for learned models.

Optimization over design spaces: mixed-integer, derivative-free, or Bayesian search.

Peer-reviewed publications in computational physics, engineering, or ML for science.

Nobody is expected to be expert in all of the above. We hire for strengths, not for the absence of weaknesses.

How We Work

We hold a high technical bar and give people full ownership of their work, from spec to ship to on-call. We write contracts before logic, test against real systems instead of mocks, and favor simple designs that ship over clever ones that do not. Our development process is AI-native: we work with agentic coding tools daily, write specs that are legible to humans and agents alike, and lead with leverage.